Home Blog

Mermaid Diagrams in ChatGPT for iOS Codex Tasks: What Is Supported

Muawia Ali
-
0
Mermaid Diagrams in ChatGPT for iOS Codex Tasks: What Is Supported

OpenAI’s wording on Mermaid diagrams in ChatGPT is narrower than many headlines suggest. The confirmed change applies to Codex task transcripts in ChatGPT for iOS. OpenAI says those transcripts can render Mermaid diagrams inline. The announcement does not establish the same behavior in an ordinary ChatGPT conversation, and it does not specifically confirm Mermaid rendering on iPad.

That distinction matters if you arrived here expecting a general diagram button on every Apple device. This article explains what OpenAI actually announced, what inline rendering changes inside a Codex task, how to make a sensible request without assuming undocumented controls, and where the evidence stops.

Mermaid diagram rendering in a ChatGPT for iOS Codex task transcript
The confirmed scope is a Codex task transcript in ChatGPT for iOS, not every ChatGPT conversation.

What OpenAI confirmed

In its July 20, 2026 product update, OpenAI listed several changes under “ChatGPT for iOS updates.” The relevant line says that support was added for rendering Mermaid diagrams inline in task transcripts. The surrounding notes refer to Codex tasks, interactive forms, task switching, hosts, workspaces, and task-list improvements. Taken together, the context identifies this as a Codex task feature inside the iOS app.

You can read the wording in the official OpenAI release notes. That is the primary source for the scope described here.

“Render inline” describes presentation. Instead of seeing only a text block that contains Mermaid notation, a user may see the resulting diagram in the task transcript. This can make a task result easier to scan on a phone, especially when the output describes a sequence, a branching decision, or relationships between components.

The note does not say that ChatGPT invents a new diagram format. Mermaid remains text-based diagram notation. The announced change concerns how that notation appears in a particular transcript context.

What a Codex task transcript is in this context

The phrase “task transcript” is important. A transcript is the record of the interaction and output associated with a Codex task. It is not a broad synonym for every chat in the ChatGPT app. OpenAI’s note places diagram rendering alongside other Codex task changes, including interactive forms and improvements to task lists and task transitions.

For the reader, the practical difference is simple. If a Codex task produces Mermaid notation in a supported iOS transcript, the app can present the diagram inline. You should not infer that pasting Mermaid text into any standard conversation will produce the same result. It might behave differently, but the cited announcement does not answer that question.

If you need a broader introduction to using Codex outside a narrow coding stereotype, see our guide to Codex for knowledge work, research, and documents. For help finding past material across ChatGPT, the ChatGPT unified search guide covers a separate product feature and should not be confused with task transcript rendering.

How to request a Mermaid diagram without overpromising

OpenAI’s release note confirms rendering, but it does not publish a required prompt formula. There is no source-backed reason to claim that one magic phrase, menu, or button is necessary. A plain request is the safest starting point. Describe the information you want organized, name the type of relationship, and ask for Mermaid notation.

For example, a request could read:

Summarize the task flow as a Mermaid flowchart. Keep the labels short, show the decision points, and include the Mermaid source in the task transcript.

That wording is an example, not an official OpenAI command. It gives the task a clear output shape while avoiding claims about controls that the release note never mentions. If the task returns a diagram inline, compare it with the surrounding transcript. The picture should reflect the actual steps and decisions described in the text.

For a sequence rather than a branching process, you can ask for a Mermaid sequence diagram. For relationships among components, ask for a compact graph. These are writing choices in your request. They are not claims that every diagram type has been separately documented or tested in the iOS app.

A careful workflow for reading the result

Inline rendering is convenient, but the diagram should not replace the transcript. Start with the task’s written result. Identify the actors, actions, and decision points that matter. Then read the diagram in the same order and check whether any label compresses the meaning too far.

  1. State the subject and the intended diagram form in the task request.
  2. Ask for short labels so the result remains readable on a phone screen.
  3. Compare each node or step with the text in the transcript.
  4. Look for missing branches, reversed arrows, or labels that imply certainty the task did not establish.
  5. Keep the text version when the diagram will inform code, documentation, or a decision.

This review is especially useful when a flow has exceptions. A neat diagram can hide an unresolved condition simply because the visual layout looks complete. The transcript gives you the context needed to notice that gap.

Do not treat the rendered view as proof that the underlying process is correct. Rendering tells you that the app can display the notation. Accuracy still depends on the task output and on your review of the content.

Where inline diagrams can help

The most natural use is a result that already has structure. A software task may describe how a request moves through services. A debugging task may return a decision tree. A repository review may explain relationships among modules. In each case, a compact diagram can reduce the effort needed to follow the result on a small screen.

Diagrams may also help when you return to a long-running task. A visual summary can act as a map of the transcript, while the text preserves details such as assumptions and unresolved questions. This is a reasonable way to use the feature, but it is not an OpenAI guarantee that Codex will always choose the right structure or preserve every detail automatically.

The feature is less useful when the result is mainly prose, contains many long labels, or changes too often for a snapshot to stay accurate. In those cases, asking for a diagram may add clutter. Use one when relationships are the point, not merely because the transcript can render it.

What the source does not establish

The release note is short, so several boundaries deserve explicit treatment.

  • Ordinary ChatGPT chats: OpenAI does not say that Mermaid diagrams render inline in every standard chat. The statement is limited to task transcripts.
  • iPad support: The same update mentions improved iPad navigation, but it does not connect that line to Mermaid rendering. An iPad-specific availability claim would go beyond the source.
  • A universal rollout: The note does not give plan eligibility, account requirements, regional availability, or a rollout percentage for this item.
  • A particular app version: No required version number is stated in the cited note.
  • Every Mermaid diagram type: The note says Mermaid diagrams in general, but it does not list or certify individual syntax types.
  • Export or editing controls: The announcement does not describe exporting the rendered image, editing it visually, copying source, or changing styles.

Those omissions are not evidence that the capabilities are unavailable. They mean only that this announcement does not confirm them. Product behavior can also change after a release note is published, so the current app and current documentation remain the final references for your account.

Checklist for evaluating Mermaid diagrams in iOS Codex task transcripts
Check the transcript, diagram labels, branches, and source scope before relying on the visual result.

Troubleshooting without inventing a hidden setting

If a task transcript shows Mermaid source as plain text, avoid assuming that a secret toggle exists. The official note does not document one. First, confirm that you are looking at a Codex task transcript in ChatGPT for iOS rather than an ordinary conversation. Then check whether the output is actually Mermaid notation and whether the task completed normally.

You can ask the task to return a smaller, simpler diagram and to include the source. That may help distinguish a complicated output from a presentation issue, but it is not a guaranteed fix. You can also consult the current product documentation or support information for account-specific behavior. Without a source, claims about reinstalling the app, changing a plan, enabling an experimental menu, or using a precise UI path would be guesswork, so this guide does not prescribe them.

How to describe the feature accurately

When sharing this update with colleagues or readers, use wording that keeps the product boundary intact:

ChatGPT for iOS can render Mermaid diagrams inline in Codex task transcripts, according to OpenAI’s July 20, 2026 release note.

Avoid saying “ChatGPT now supports Mermaid everywhere” or “Mermaid works on iPhone and iPad chats.” Neither statement follows from the source. Accurate wording may sound less dramatic, but it helps users understand where to look and prevents a task-specific update from becoming a false platform-wide promise.

FAQ

Can every ChatGPT chat on iPhone render Mermaid diagrams?

The cited OpenAI note does not confirm that. It specifically says Mermaid diagrams render inline in task transcripts, and the surrounding context is about Codex tasks in ChatGPT for iOS.

Does the announcement confirm Mermaid diagrams on iPad?

No. The update separately mentions an iPad navigation improvement, but it does not explicitly say that Mermaid diagram rendering is supported on iPad. The existing page URL is preserved, but the article title and claims have been corrected to match the source.

Do I need a special Mermaid prompt?

OpenAI does not provide a required prompt in the release note. You can make a plain-language request for a Mermaid flowchart, sequence diagram, or graph, then review the returned source and visual against the transcript.

Can I export or edit the rendered diagram?

The cited announcement does not document export or visual editing controls. Check the current app and official documentation for any later additions rather than assuming those tools are part of the July update.

The useful takeaway

The confirmed feature is modest and specific: Mermaid diagrams can render inline in ChatGPT for iOS Codex task transcripts. That can make structured task output easier to read on a phone. It does not, by itself, establish Mermaid rendering in ordinary chats, iPad support, export tools, account eligibility, or complete coverage of Mermaid syntax. Ask for a clear diagram when the task benefits from one, keep the source and transcript nearby, and judge the visual by whether it accurately represents the work.

Related update: Apple Intelligence Expands to iPad and Mac.

OpenAI Codex for Knowledge Work: Research, Documents, and No-Code Workflows

Muawia Ali
-
0
OpenAI Codex for Knowledge Work: Research, Documents, and No-Code Workflows

OpenAI Codex began as a software development tool, but OpenAI now describes a much wider pattern of use. Analysts, marketers, operators, designers, researchers, investors, and bankers are using Codex for research, data analysis, documents, presentations, spreadsheets, and small internal tools. That makes OpenAI Codex for knowledge work a more useful subject than the vague claim that AI will “transform work.” We can look at the jobs people are handing to it, the outputs it creates, and the limits of the evidence.

OpenAI reported on June 2, 2026 that Codex had more than 5 million weekly active users. Knowledge workers accounted for about 20 percent of users and were growing more than three times as fast as developers. Those figures come from OpenAI, not an independent usage audit, but they show how the company sees the product changing: Codex is being positioned as a workspace for making things, not simply as a place to ask questions.

OpenAI Codex knowledge work flow from connected sources to research, documents, and review
Codex can connect source material to research and work products, but the final review still belongs to the user.

What knowledge work means in Codex

In OpenAI’s account, knowledge work includes creating reports, spreadsheets, presentations, contracts, and other finished materials. Research, data analysis, workflow automation, and lightweight tool building are also prominent. The fastest-growing task groups among knowledge workers were data analysis, research, and the creation of what OpenAI calls knowledge artifacts.

The distinction between an answer and an artifact matters. A chat response may explain a sales trend. A work product could be the spreadsheet that calculates it, the chart used in a meeting, the written summary, and a small dashboard where colleagues can explore the assumptions. Codex is intended to work across more of that chain.

OpenAI also says users increasingly run several Codex tasks in parallel. A researcher might investigate separate questions while another task organizes findings. An operator could prepare a report while a second task turns the same context into tickets. Parallel work may reduce waiting, but it does not prove that every result is correct or that the tasks agree with one another. More output can also mean more material to inspect.

Research and information spread across tools

Finding the right information is often harder than writing the final memo. Notes may be split between chat threads, shared documents, project trackers, and data systems. OpenAI says Codex can work with connected tools and organizational context through plugins. The practical promise is that a task can begin with the material where it already lives instead of requiring someone to collect every excerpt by hand.

Zapier offers a concrete example in OpenAI’s announcement. Its teams use Codex to pull knowledge from Slack, Google Docs, and Coda, then turn that context into postmortems, incident response plans, and feature tickets. This is more specific than saying Codex “does research.” The system gathers relevant internal material and reshapes it for a known business purpose.

That example does not establish that Codex will find every relevant message, resolve conflicting records, or judge which source is authoritative. Access to a tool is not the same as complete understanding of it. If a postmortem depends on dates, owners, or causal claims, a person familiar with the incident still needs to check the evidence. Readers who deal with information scattered across ChatGPT itself may also find our ChatGPT unified search guide useful.

Documents, spreadsheets, slides, and annotations

Codex is not presented only as a research collector. OpenAI says nontechnical teams inside the company use it to prepare executive materials, create dashboards, build internal apps, and turn creative briefs into work that follows brand and design constraints. These are production tasks with different kinds of structure. A spreadsheet needs formulas and traceable inputs. A presentation needs a readable argument. A branded asset must follow visual constraints rather than merely contain the right words.

Annotations are meant for the revision stage. A user can select a particular part of a document, spreadsheet, slide, site, Markdown file, or code file and request a focused change. OpenAI gives several examples: select a site’s navigation bar and ask for a different font, highlight a claim in an investment thesis and ask where it came from, or mark a chart and request a clearer label.

This is a sensible editing model because useful work rarely arrives in one perfect draft. A focused instruction can preserve the sections that already work. Still, the source describes how annotations direct a revision, not a guarantee that the revised claim, formula, or chart label is accurate. The user must read the changed section and check whether the surrounding material still makes sense.

Plugins and what “no code” means here

OpenAI introduced role-specific plugins that package tools, context, and workflows for a particular kind of work. Its announcement names a data analytics plugin for business performance analysis and says six role-specific plugins were launching. It also listed Corporate Finance, Private Equity Investing, Marketing Strategy, Strategy Consulting, and Legal among the plugins planned to come later.

The phrase “no coding required” refers to using the new role-specific plugins, according to OpenAI. It should not be stretched into a claim that every workflow is effortless or that technical work disappears. Teams can adapt plugins, build custom plugins, and share them for their own systems. Custom connections and specialized processes may still require people who understand the underlying data, permissions, and expected output.

A useful way to think about a plugin is as a prepared route through a task. It can give Codex access to relevant tools and establish a repeatable method. It cannot supply missing records, correct a bad source system, or decide what an ambiguous business term means for a particular company. If “active customer” has three definitions in three departments, connecting another tool will not settle the disagreement.

From a model to an interactive site

OpenAI also announced Sites, a preview feature for Business and Enterprise customers. The company describes Sites as hosted, interactive websites and apps that can be shared with people in a workspace by URL. Examples include dashboards, planners, review workspaces, project boards, galleries, and other lightweight tools.

The examples are concrete. Codex could turn material for a customer review into a page containing product updates, open questions, usage trends, and next steps. It could convert a financial model into a scenario planner where leaders compare assumptions. Launch materials could become a shared hub for current messaging, milestones, owners, and decisions. OpenAI says a site can be updated as those details change.

These examples explain the intended shape of the feature. They do not tell us how well any specific site will handle complex permissions, regulated data, heavy traffic, unusual calculations, or long-term maintenance. Nor do they show that a generated site is a replacement for a production application. “Lightweight tools” is the company’s own framing and is a useful boundary. For related coverage, see our guide to building and publishing sites with ChatGPT.

Documented examples from OpenAI, Zapier, and NVIDIA

The strongest part of OpenAI’s announcement is the set of named examples. Inside OpenAI, nontechnical teams make internal apps, dashboards, executive materials, and creative work. Zapier uses connected company knowledge for incident and product documents. At NVIDIA, researchers use Codex across experiment workflows, including finding research ideas and writing scripts for machine learning infrastructure.

These examples cover different handoffs. The OpenAI case moves from a brief or business need to a deliverable. The Zapier case moves from scattered internal context to structured records and plans. The NVIDIA case links scientific exploration with the scripts needed to support experiments. None of the examples says that people leave Codex alone to approve an executive report, determine the cause of an incident, or validate a scientific result.

OpenAI’s separate knowledge-work report also says people use Codex to find information buried across systems, coordinate work across tools and teams, produce deliverables, and move projects through review and approval. The final phrase is worth noticing. Review and approval remain part of the process. Codex may prepare and route work, while accountable people still decide whether it is ready.

Review questions for Codex research, documents, data analysis, and lightweight sites
Check the sources, calculations, permissions, and final output before using Codex work in a decision.

What can be concluded, and what cannot

The official sources support a narrow set of conclusions. Codex is used by developers and nondevelopers. It can create several common work products, use connected context through plugins, accept targeted annotations, run tasks in parallel, and create shareable lightweight sites in the preview described by OpenAI. Named organizations report using it for real internal workflows.

The sources do not provide an independent comparison with competing products. They do not establish an error rate for research, financial analysis, contracts, spreadsheets, or generated sites. They do not claim that every plugin works with every system, that every role-specific plugin is already available, or that a user can safely skip subject-matter review. This article is based on the published material and does not present personal testing.

That leaves a straightforward standard for readers. Judge a Codex workflow by the evidence visible in its output. Can you trace a claim to its source? Do spreadsheet totals reconcile? Are assumptions shown? Does a generated site expose only the intended information? Can a reviewer understand what changed after an annotation? Those questions test the work itself without pretending that a product announcement answers them.

Official sources

FAQ

Is Codex only for software developers?

No. Developers remain its largest user group, but OpenAI says knowledge workers account for about 20 percent of Codex users. The company specifically names analysts, marketers, operators, designers, researchers, investors, and bankers.

Can Codex create documents and spreadsheets?

OpenAI says knowledge workers use Codex to create reports, spreadsheets, presentations, contracts, and other work products. Annotations can target a selected part of documents, spreadsheets, and slides for revision. The sources do not guarantee the factual or mathematical accuracy of every output.

Do Codex plugins require coding?

OpenAI says its role-specific plugins are designed for use with no coding required. Teams can also customize plugins or build their own, which may call for technical knowledge depending on the systems and process involved.

Does Codex replace human review?

The official announcements do not say that it does. Their examples include work moving through review and approval, and annotations are built around human feedback. Important claims, calculations, permissions, and decisions still need an accountable reviewer.

ChatGPT Health Guide: Connect Apple Health and Medical Records

Muawia Ali
-
0
ChatGPT Health Guide: Connect Apple Health and Medical Records

ChatGPT Health can connect selected health information to your conversations, but the feature is not a replacement for a clinician or a shortcut to a diagnosis. Its practical value is narrower and more useful: it can help you read your own records in context, compare results over time, prepare questions for an appointment, and relate Apple Health activity or sleep data to the routine you describe.

This ChatGPT Health guide explains the current U.S. rollout documented by OpenAI on July 23, 2026. It covers eligibility, supported connections, permission prompts, privacy controls, and the limits that matter before you connect anything. OpenAI’s January announcement described an earlier waitlist and a more separate Health experience. The newer rollout changes how connected context can be used, so this guide relies primarily on the July announcement.

Medical disclaimer: ChatGPT can make mistakes. It does not replace the care or judgment of a qualified medical professional and should not be used for diagnosis or treatment. Verify important information against the original record and discuss medical decisions, symptoms, medications, and treatment choices with your healthcare provider. Seek appropriate professional or emergency care when needed.

Who can use ChatGPT Health now?

OpenAI says Health is rolling out to logged-in ChatGPT users who are 18 or older in the United States. The rollout covers the web and iOS versions of ChatGPT and includes Free, Go, Plus, and Pro plans. Health is not available in Codex.

Those details are easy to blur together, so it helps to separate access to Health from access to a particular data source:

  • The ChatGPT Health experience is rolling out to eligible U.S. adults on web and iOS.
  • Connecting Apple Health requires iOS.
  • Medical record integrations are limited to supported U.S. hospital systems and named supported services.
  • Availability may still depend on the rollout reaching your account.

You can still ask health questions in ChatGPT without connecting a record, account, or app. A connection is optional. Its purpose is to let ChatGPT consider information you have chosen to sync rather than making you paste the same lab values, medication list, or activity details into every conversation.

What you can connect

The current OpenAI documentation lists Apple Health, supported medical records from U.S. hospital systems, One Medical, and Function Health. Apple Health can also act as a bridge for information from a wearable, fitness app, or nutrition app that already shares data with Apple Health.

That bridge does not mean every metric will transfer. OpenAI notes that available information varies by app and metric, and some proprietary scores may not be shared through Apple Health. If a wearable app is involved, its sharing setting must allow the relevant information to reach Apple Health before ChatGPT can use what Apple Health makes available.

ChatGPT Health connection flow for Apple Health and supported medical records
Connect only the source you want, let it sync, then review the information before relying on it in a conversation.

How to connect Apple Health or medical records

OpenAI’s current instructions begin in Health, which appears in the ChatGPT sidebar or More menu. Select Get started and follow the guided connection process. Once a source is connected, syncing may take a few minutes. You can connect another account while it syncs or return later.

For Apple Health, use an iOS device and confirm that any wearable or app you care about is already sharing the relevant data with Apple Health. For medical records, choose from the supported providers presented to your account. Do not assume that every hospital portal in the United States is supported merely because medical record connections exist.

After syncing, review conditions and medications. OpenAI says you can remove items that are no longer relevant and add missing context, including family health history. This review is not clerical busywork. A connected record may be incomplete or out of date. A medication, for example, may remain listed after you have stopped taking it. Check important details against the original source and tell ChatGPT when something has changed.

How permission works in conversations

By default, ChatGPT asks before using connected medical records or Apple Health information to personalize a response. You can approve access for that request, or choose to always allow access. Always allowing access turns off those permission prompts. OpenAI says you can change this choice at any time in Settings > Plugins > Health.

You can also add @Health to a message when you want to ask explicitly for connected health context. Once information is synced and permission is available, a question does not have to begin inside the Health area. ChatGPT may use relevant Health information elsewhere in ChatGPT according to the permissions you selected.

This is one of the main differences between the initial January version and the July rollout. Early testing required people to visit a separate Health space for the full benefit. OpenAI later reported that more than 70 percent of health-related conversations among people with access happened outside that space. Health remains the home base for connections, records, recent trends, prompts, and past health conversations, but connected context can now be brought into other conversations with permission.

If you want more control, leave the default request-by-request permission in place. It gives you a chance to decide whether connected information belongs in the current answer. If you select always allow, remember that you can return to the Health plugin setting and change it.

What OpenAI says about privacy and data use

OpenAI states that all ChatGPT conversations are encrypted at rest and in transit. Information connected in Health receives additional encryption protections. The company also states that connected medical records, Apple Health information, and conversations that use that information are not used to train its foundation models or target ads. This applies regardless of the model-training setting selected for ChatGPT.

That statement is specific. Conversations that do not use connected Health data continue to follow your normal ChatGPT model-training choice, which OpenAI says you can change in Settings > Data Controls. It is better to keep this distinction than to turn the policy into a broad claim that no ChatGPT conversation is ever used for training.

You can disconnect a source in Health > Accounts. According to OpenAI, data synced from that source is deleted from its systems within 30 days after disconnection. Information already included in your ChatGPT conversation history remains there until you delete those conversations. Disconnecting a source and deleting prior chats are therefore separate actions.

Memory is another separate control. OpenAI says memories can be created from Health conversations, but they are not created directly from connected medical records or Apple Health information. To have a conversation without creating memories, use Temporary Chat or turn memory off in Settings > Memory.

OpenAI also describes extra safeguards when an action in another connected plugin could disclose Health information. Its example is sending a running partner a training plan based on Apple Health data. Some sensitive actions may require confirmation. This does not justify an absolute privacy guarantee, so review the destination and the information being shared whenever ChatGPT asks you to confirm an external action.

Useful questions that stay within the feature’s limits

Connected context is most helpful when the task is about understanding, organizing, or preparing. You might ask ChatGPT to summarize changes between recent and earlier lab results, explain unfamiliar terms in a visit note, list questions for a follow-up appointment, or compare sleep and activity patterns with a routine you describe.

Use clear prompts and ask the answer to identify uncertainty. If a result looks wrong, compare it with the original portal or report rather than asking the model to resolve the discrepancy by guessing. Correct information that is incomplete or stale. For broader prompt habits, our guide to ChatGPT custom instructions explains how persistent preferences differ from details supplied for one task. You may also want to review the site’s guide to ChatGPT memory and its controls before deciding how much personal context should persist.

A cautious prompt can be simple: “Using the lab results I allowed you to access, summarize what changed since the prior test, quote the dates and values you relied on, flag missing context, and help me prepare questions for my clinician. Do not diagnose me or recommend changing medication.” This frames the job as preparation and asks ChatGPT to show its basis.

A short safety and privacy checklist

Privacy and medical safety checklist for using ChatGPT Health
Check source accuracy, permissions, memory, and conversation history before treating connected context as ready to use.
  1. Confirm that your account is eligible and that the connection you need is supported.
  2. Connect only the source you intend to use.
  3. Review synced medications, conditions, dates, and values against the original source.
  4. Keep the default permission prompt if you want to approve each use of connected data.
  5. Check memory and Data Controls separately because they govern different behavior.
  6. Disconnect a source when you no longer want it synced, then delete relevant conversations if you also want that history removed.
  7. Do not use ChatGPT for diagnosis or treatment decisions. Verify important information and speak with a qualified healthcare professional.

Where ChatGPT Health may help, and where it stops

Health can reduce the effort of gathering scattered context for a question. It can help turn a long visit note into plain language, arrange a timeline, notice that two test dates differ, or help you draft a focused list for an appointment. Those are information tasks. They can make a clinical conversation easier without transferring clinical responsibility to the model.

The limits remain substantial. ChatGPT can make mistakes. Synced information may be missing or old. A wearable may expose only some metrics, and a hospital system may not be supported. Even a correct summary does not account for a physical examination, full medical history, or the judgment of a professional who is responsible for your care.

Use connected Health information as material to review, not as a verdict. If a response affects an important decision, verify the underlying record and take the question to your provider. For urgent symptoms or emergencies, seek appropriate professional help rather than waiting for a chatbot response.

Frequently asked questions

Is ChatGPT Health available outside the United States?

OpenAI’s July 23, 2026 rollout announcement says Health is rolling out to logged-in users age 18 and older in the United States. It does not announce general availability outside the United States in that release.

Can I connect Apple Health from the web?

The Health experience is rolling out on web and iOS, but OpenAI specifically says that connecting Apple Health requires iOS. After information is connected and synced, its use remains subject to your Health permissions.

Does OpenAI train its foundation models on connected health information?

OpenAI says connected medical records, Apple Health information, and conversations that use that information are not used to train its foundation models or target ads. Conversations that do not use connected Health data follow your standard model-training setting.

What happens after I disconnect a health account?

OpenAI says data synced from that source is deleted from its systems within 30 days. Information already included in your ChatGPT conversation history remains until you delete those conversations.

Official sources

Related update: Apple Intelligence Expands to iPad and Mac.

ChatGPT Unified Search Guide: Find Chats, Projects, Images, and Documents

Muawia Ali
-
0
ChatGPT Unified Search Guide: Find Chats, Projects, Images, and Documents

Finding an old ChatGPT item used to mean remembering where it lived. Was it a normal conversation, a Project, an uploaded document, or an image buried in a thread? ChatGPT unified search reduces that hunting. OpenAI announced on July 14, 2026 that one search experience can find chats, projects, images, and documents, with results available from the ChatGPT sidebar.

This guide separates OpenAI’s documented facts from practical search advice. The release note confirms the content types, platforms, filters, result behavior, and availability, but it does not explain every ranking rule.

What ChatGPT unified search is

ChatGPT unified search is a single place for searching across several kinds of content associated with your ChatGPT use. According to the official ChatGPT release notes, it searches past chats, projects, images, and documents. OpenAI says you start from the sidebar, can narrow the results by content type, and can select a result to open the relevant chat, project, or file directly in ChatGPT.

The word “unified” describes the retrieval surface. You do not have to begin in one search box for conversations and another for Project material. A single search can surface different kinds of saved work, while content type filters help when the first result list is too broad.

This is retrieval, not web search. It is meant to locate material in ChatGPT, not to find current information on the public internet. It also is not the same as asking ChatGPT to answer a question from memory. Search returns items that you can open and inspect.

Workflow for searching ChatGPT chats, projects, images, and documents from one sidebar
Begin with a distinctive phrase, review the mixed results, then narrow by content type if needed.

Where the new search is available

OpenAI says the updated search rolled out on the ChatGPT web app, iOS, and Android. The same release note states that it is available globally on all ChatGPT plans. That is stronger than a limited beta or a paid-only rollout.

The July 14 platform list does not mention the desktop app. A separate July 16 note describes unified Recents on desktop, which is not the same as confirming unified search there. For the documented experience, use the web, iOS, or Android app.

Interface placement can differ slightly between a wide browser window and a phone screen. The July 14 note keeps the instruction simple: start from the ChatGPT sidebar.

What you can find

Past chats

Chats include conversations you previously had in ChatGPT. The release note says past chats can appear in the unified results. Selecting a chat result opens that conversation directly, so you can inspect its context and continue working if appropriate.

Projects

Projects keep related chats and source material together. Unified search can return a Project itself, which is useful when you remember the subject but not the particular conversation that contains the detail. For more on how Project material is organized, see our guide to organizing research with ChatGPT Project sources.

Images

The release note explicitly includes images among searchable content. This can help when the item you remember is visual rather than a memorable chat title. For example, you may recall a diagram, a draft graphic, or the name used when discussing an image. OpenAI does not explain in the release note whether matching comes from filenames, surrounding conversation text, image content, or a combination. Treat any suggested query as an experiment, not a guaranteed image recognition command.

Documents and files

OpenAI says the search covers documents and that selecting a file result opens it directly in ChatGPT. That makes unified search useful for locating an uploaded brief, report, spreadsheet, or other document without first finding the conversation that contained it. The source does not publish a list of searchable file formats or explain whether every word inside every format is indexed.

How to use ChatGPT unified search

  1. Open ChatGPT on the web, iOS, or Android and sign in to the account or workspace where the item was used.
  2. Open the sidebar and choose Search.
  3. Enter a distinctive word or phrase that you remember from the item.
  4. Review the results across chats, projects, images, and documents.
  5. If the list is crowded, use a content type filter. The release note confirms content type filtering, though it does not list every label or control shown in each app.
  6. Select the result. ChatGPT opens the matching chat, project, or file directly.

Opening a result is important. A search match tells you that an item may be relevant; it does not replace checking the surrounding conversation, the document version, or the Project context. If the result is an old chat, review its date and nearby messages before reusing a decision or instruction.

Practical search strategies

The ideas in this section are search advice, not official query syntax. OpenAI has not documented special operators, Boolean commands, date ranges, or ranking controls for unified search in the cited release note.

Start with the least common phrase you remember

A generic query such as “meeting” may match many items. A product codename, client-approved phrase, unusual error message, document heading, or exact sentence fragment is more likely to reduce the list. If the exact phrase fails, try one distinctive word from it, then a second query with another remembered term.

Use the content type filter after the first search

Searching broadly first lets you see whether the item was stored somewhere other than expected. If the result list is noisy, narrow it. Choose chats when you remember a discussion, Projects when you want the container for ongoing work, images when the visual is the target, or documents when you need an uploaded file.

Search for durable language, not temporary context

People often remember why they needed an item but not the words they used. Search works better when you recall text likely to be present, such as a heading, filename, quoted requirement, error code, or named deliverable. “The budget conversation from last month” describes your memory. “Q3 regional budget assumptions” is more likely to resemble stored text.

Try the title and body separately

For a chat, first search a term likely to be in its title. If that does not help, search an exact phrase from the conversation. For a file, try the filename, then a heading or phrase you remember discussing near the upload. These are ordinary retrieval tactics, not promises about how the search index treats each field.

Confirm the account and workspace

If a query returns nothing, check that you are using the account or workspace where the work occurred. Search cannot be assumed to cross separate personal and organizational histories. The official release note does not claim cross-account or cross-workspace retrieval.

Checklist for refining a ChatGPT unified search and confirming the opened result
Use a specific phrase, narrow the content type, open the result, and verify its context before relying on it.

A simple recovery workflow for hard-to-find items

If your first query misses, do not keep adding words to one long sentence. Use a short sequence of focused attempts:

  1. Search the exact phrase or filename you remember.
  2. Remove common words and keep the rarest term.
  3. Filter to the most likely content type.
  4. Try a second exact term that probably appears in the item.
  5. Open plausible results and check the date, surrounding messages, Project name, or file context.

This workflow avoids pretending that unified search has undocumented advanced syntax. It uses the two controls OpenAI does confirm: a search entry point and content type filters. It also accounts for ordinary memory errors, such as remembering a document as a chat or assuming an image lived in a Project when it was attached elsewhere.

For people who do much of their work inside Projects, better organization still helps. Search is a recovery tool, not a substitute for clear Project names and relevant source material. Our ChatGPT Projects guide explains the broader Project workflow.

History and privacy caveats

Search results necessarily depend on material available to the account or workspace you are using. In another July 2026 note about conversation history, OpenAI says that history remains available subject to the user’s plan, workspace settings, and account access. The unified search announcement does not claim that search bypasses any of those boundaries.

Unified search does not change the need to follow your organization’s data rules. Do not upload confidential material merely to make it easier to find later. The July 14 announcement does not describe new privacy settings, retention periods, deletion recovery, administrator visibility, or training controls for unified search. Check the controls and policies that apply to your account or managed workspace.

What the official sources do not establish

The announcement is concise, so several details remain undocumented in the sources cited here:

  • OpenAI does not publish the ranking formula or explain how results are ordered.
  • The release note does not promise semantic matching, typo tolerance, fuzzy search, Boolean operators, quotation operators, or date filters.
  • It does not list supported document formats, indexing size limits, or whether text inside every uploaded file can be searched.
  • It does not explain how image matches are generated.
  • It does not confirm search across separate accounts or workspaces.
  • It does not say that search can restore deleted items.

These gaps are not proof that a capability is absent. They only mean the cited official material does not establish it. Interfaces can also change after a rollout, so consult OpenAI’s release notes if the location or available controls differ from this guide.

Frequently asked questions

What does ChatGPT unified search search?

OpenAI says it searches chats, projects, images, and documents from one place in ChatGPT.

Which plans and platforms support unified search?

The July 14, 2026 release note says it is available globally on all ChatGPT plans on the web, iOS, and Android.

Can I filter unified search results?

Yes. OpenAI says you can narrow results by content type. The release note does not document every filter label or any advanced query operators.

Does the announcement promise recovery of deleted chats?

No. The unified search release note does not say that deleted chats can be restored through search. Treat deletion and retention as separate account or workspace controls.

The useful takeaway

ChatGPT unified search gives you one sidebar entry point for several kinds of saved work. Begin with specific remembered wording, inspect the mixed results, narrow by content type when needed, and open the item to verify its context. Those steps stay within what OpenAI has documented while making the feature useful for everyday retrieval.

ChatGPT Sites Guide: Build, Review, and Publish a Hosted Site

Muawia Ali
-
0
ChatGPT Sites Guide: Build, Review, and Publish a Hosted Site

ChatGPT Sites turns a prompt or compatible project into a hosted website, web app, dashboard, or game. The building and publishing work happens through Sites in ChatGPT on the web or in the desktop app, with Codex handling the creation and revision process. This is not a general promise of unrestricted consumer web hosting. Sites is in public beta, availability depends on your plan, region, and workspace settings, and every account is subject to plan-specific usage limits.

The feature is best understood as a short path from an idea to a focused hosted experience. You describe the audience and behavior, review what ChatGPT creates, request changes, save a version, and deploy only after approval. OpenAI provides the hosting layer and supports access controls, environment values, analytics, and optional data services within the Sites runtime.

What ChatGPT Sites is

According to the official Sites documentation, Sites can create, host, refine, and share websites, web apps, and games. A Site remains available in the Sites list after the chat that created it ends. It is also separate from a ChatGPT Project, even when project files or a compatible local codebase helped produce it.

OpenAI first presented Sites on June 2, 2026 as a preview for Business and Enterprise teams. The current documentation calls it a public beta and says access may vary by plan, region, and workspace configuration. That distinction matters. Check the Sites entry in your own ChatGPT account rather than assuming that another user’s access or limits apply to you.

Sites has management views in ChatGPT on the web and desktop. It does not have a standalone Sites management view in Codex CLI or the IDE extension. You can still use those tools to edit and test a local project, but creating, saving, deploying, and managing the hosted Site happens in ChatGPT.

If you are comparing this with broader knowledge-work features, our guide to ChatGPT Work outputs explains how websites fit beside documents, presentations, and spreadsheets. The separate guide to Codex for knowledge work covers the wider research and production workflow.

Choose a small job before you build

A strong first Site has one audience and one recurring task. OpenAI’s Academy examples include a launch tracker, weekly review dashboard, option calculator, onboarding page, and internal project page. Each is narrow enough to review without pretending that a generated first draft is a finished software product.

A launch tracker might contain milestones, owners, links, risks, and weekly updates. A calculator could accept a few inputs, state its assumptions, and show how it reached the result. An onboarding page could collect the project background, useful documents, team contacts, common questions, and a first-week checklist. These are easier to inspect than a large application with many roles, integrations, and hidden business rules.

Write a brief before opening Sites. Name the people who will use it, the decision or task it supports, the information it needs, what users can change, and what must remain private. If the Site needs to remember records or accept uploads, say that explicitly. If a simple content page is enough, do not request storage merely because the platform offers it.

ChatGPT Sites workflow from scoped brief through review, saved version, deployment, and sharing

A practical build, review, and update workflow

1. Describe the audience and required behavior

Open Sites from the ChatGPT desktop app, or choose More > Sites on the web. You can also go to chatgpt.com/sites. In your request, use the word “website” or mention @Sites to start the workflow explicitly.

A useful request is concrete: “Build a website for our weekly product launch review. Show milestones, owners, due dates, current risks, and links to approved documents. Let reviewers filter milestones by owner. Keep access limited while I review it.” This defines the audience, page purpose, information, interaction, and initial access without dictating an unnecessary technical stack.

For an existing local project, ask ChatGPT to check compatibility before deployment. Sites runs within a supported runtime, so some frameworks, private networks, databases, background services, and hosting patterns will not work. Compatibility should be established rather than assumed.

2. Inspect the first version as a visitor

Review the generated page before discussing polish. Confirm that the labels match your workflow, links point to the intended resources, forms collect only necessary information, and calculations use the stated rules. Try the main path from the perspective of a new visitor. A clear design cannot compensate for a wrong formula, a misleading status, or access that is broader than intended.

OpenAI’s Academy guide recommends checking content, structure, labels, calculations, and interactive elements. The current product documentation also allows edits from the preview: choose Edit, describe the change, and attach a screenshot or relevant file when visual context would help.

3. Request precise revisions

Make revision requests observable. Instead of asking for a “better dashboard,” ask to move unresolved risks above completed milestones, add an owner filter, explain the scoring formula beside the result, or shorten the onboarding checklist. Reopen the preview after each meaningful change and repeat the core user path.

Keep factual and editorial review separate from visual review. First check the source material, numbers, permissions, and behavior. Then review spacing, hierarchy, mobile layout, empty states, error messages, and keyboard use. If the Site came from a local project, inspect source changes and database migrations in the Codex review pane before publishing.

4. Save a version before deployment

Sites separates saving from deploying. Saving creates a reviewable, deployable version. Deploying publishes a saved version and returns its production URL. Every Sites deployment URL is a production deployment, so a deploy action should never be used as a private preview shortcut.

Ask ChatGPT to save a version without deploying it, review that candidate, and deploy only the approved version. You can also ask it to list or inspect saved versions if several candidates exist. This small release discipline makes later updates easier to reason about.

5. Set the narrowest suitable audience

A new Site is limited to its owner and workspace admins until access changes. Depending on the account and workspace, choices may include selected active users or groups, anyone in the workspace, or anyone on the internet. Internet access appears only when public publishing is enabled. Enterprise public publishing is off by default and requires an admin to enable it.

Sharing permission allows people to visit; it does not give them editing rights. Invited users in a restricted audience must sign in with the account that received access. A public Site can be visited without workspace access, but that option should not be confused with universal availability on every plan or workspace.

6. Verify the production version and maintain it

Open the production URL as the intended audience would. Test sign-in, navigation, forms, saved data, uploads, calculations, and responsive layouts. Confirm that a restricted visitor cannot reach content outside the chosen audience. Sites records traffic automatically and provides unique visitor and page-view metrics over time, although the official documentation says this analytics view is currently unavailable for Enterprise-owned Sites.

For updates, reopen the Site, describe the change, review the new result, save another version, and deploy that approved version. The OpenAI Academy page notes that Sites cannot connect directly to live data today. For frequently changing information, it suggests using a separate automation to gather updates, then reviewing and refreshing the Site. Do not describe a scheduled summary as a direct live-data connection.

Hosting, access controls, storage, and databases

OpenAI documents hosting, access controls, storage, and database support, but each belongs to the supported Sites runtime. Content-led pages can avoid persistent application state. Durable structured records, user progress, and scores can use D1, a relational database. File uploads such as images, documents, audio, and video can use R2 object storage. A searchable upload system may use D1 for metadata and R2 for the file contents.

Request durable storage only for information people expect the Site to remember. A temporary theme choice or dismissed banner does not need the same treatment as submitted project requests or saved progress. For identity-aware internal tools, Sites can use the current workspace user’s identity. Public Sites may offer optional Sign in with ChatGPT while remaining open to signed-out visitors, if that behavior is deliberately built into the Site.

Hosted environment values and secrets are managed in the Site’s settings. OpenAI says not to place secrets in prompts, attached files, Site content, or .openai/hosting.json. After changing hosted environment values, redeploy the approved saved version so the new deployment uses them.

Limits and security checks before sharing

Security and release checklist for reviewing a ChatGPT Site before sharing

During the public beta, plan-specific limits apply across all Sites. ChatGPT displays current limits and warns as you approach one. Reaching a limit can stop you from creating another Site, adding storage, or keeping a high-usage Site public, while still allowing you to edit and manage existing Sites. The official page does not publish one universal allowance, so the limit shown in your account is the relevant one.

Review generated text and images, external links, uploaded files, forms, and interactions. Remove confidential information, credentials, and third-party material you do not have permission to share. If a feature collects personal data, decide whether collecting it is necessary, explain how identity information is used, and meet applicable privacy requirements.

Sites does not support data residency or inference residency at launch. OpenAI says this covers deployed Sites, source code, D1 and R2 data, generated artifacts, and logs. Its documentation also prohibits using Sites for Protected Health Information, payment-card data, financial transactions, malware, phishing, impersonation, or experiences aimed at children under 13 or the applicable digital-consent age.

Custom domains may be available for some accounts. You must own the domain and control its DNS because Sites does not register domains. OpenAI also states that custom domains are unavailable in Enterprise workspaces at launch. Treat both public publishing and custom domains as account-dependent settings, not guaranteed parts of every Sites plan.

FAQ

Do I need to know how to code to use ChatGPT Sites?

No coding knowledge is required to start from a plain-language description. You still need to review the generated content, behavior, data handling, and access settings. More complex products may require engineering work outside Sites.

Can anyone on the internet open a ChatGPT Site?

Not by default. A new Site is limited to its owner and workspace admins. Public access is available only when the account and workspace support it, and Enterprise admins must enable public publishing.

Can a Site store records and uploaded files?

Yes, within the supported runtime. OpenAI documents D1 for durable structured data and R2 for file storage. Ask for these only when the experience needs persistent records or uploads.

Can I update a Site after it has been deployed?

Yes. Reopen it, request and review the changes, save a new version, then deploy the approved version. Because Sites cannot connect directly to live data today, regularly changing content may require a separate automation and a reviewed refresh.

Official references

This guide is based on OpenAI’s published documentation and announcements. It does not claim hands-on testing of account-specific beta access, limits, or deployment behavior.

GPT-5.6 Sol vs Terra vs Luna: Which Model to Use

Muawia Ali
-
0
GPT-5.6 Sol vs Terra vs Luna: Which Model to Use

GPT-5.6 Sol vs Terra vs Luna is a choice between three capability tiers, not three unrelated products. OpenAI describes Sol as its flagship, Terra as the balanced option for everyday work, and Luna as the fastest and most affordable member of the family. All three belong to the same GPT-5.6 generation, but they are priced and positioned for different workloads.

This comparison uses OpenAI’s launch announcement, its July 30, 2026 pricing update, and the GPT-5.6 system card. It does not rely on personal testing, and it does not assume that a company benchmark predicts the result of your own prompt. The useful question is simpler: how much capability does a task need, how often will it run, and how expensive would an error be?

GPT-5.6 family at a glance

Model OpenAI’s position Practical starting point API price from July 30, 2026
GPT-5.6 Sol Flagship and highest-capability tier Difficult, ambiguous, or high-stakes work where extra reasoning can change the outcome $5 per 1M input tokens and $30 per 1M output tokens
GPT-5.6 Terra Balanced model for everyday work General knowledge work, coding, and agent tasks that need a middle ground $2 per 1M input tokens and $12 per 1M output tokens
GPT-5.6 Luna Fastest and most affordable tier High-volume, well-specified tasks and routine steps in larger workflows $0.20 per 1M input tokens and $1.20 per 1M output tokens

The Sol price above remained unchanged in OpenAI’s July 30 update. Terra became 20 percent cheaper, while Luna became 80 percent cheaper. Those are API token prices, not ChatGPT subscription prices. OpenAI said ChatGPT and Codex subscription prices and quota budgets did not change, although Terra and Luna began consuming fewer credits on those paid products.

GPT-5.6 Sol, Terra, and Luna task selection workflow
Start with the task’s stakes, ambiguity, volume, and cost of review before choosing a GPT-5.6 tier.

What GPT-5.6 Sol is for

Sol is the family member OpenAI presents as its strongest model. The launch material emphasizes coding, professional knowledge work, computer use, cybersecurity, scientific work, and design judgment. It also supports the broadest investment in reasoning. The max setting gives the model more time to explore and check an answer, while ultra coordinates four agents in parallel by default for demanding work. OpenAI says developers can build comparable multi-agent patterns through a beta in the Responses API.

That positioning makes Sol the sensible first candidate when the task is poorly specified, crosses several tools, requires a difficult plan, or carries a large cost if the answer is wrong. Examples include diagnosing a complex software failure, reconciling conflicting source documents, or planning a multi-stage analysis. Sol can also be used for ordinary prompts, but paying for the top tier is hard to justify when a smaller model produces an acceptable result after the same review.

Speed is a separate buying decision. On July 30, OpenAI introduced Fast mode for Sol in the API, replacing Priority Processing. The company says Fast mode can deliver up to 2.5 times the speed of Standard processing at twice the price, without changing intelligence. Existing requests tagged as priority continue to work. This is useful when latency matters enough to justify a processing premium, but it is not a new model tier.

What GPT-5.6 Terra is for

Terra is the middle option. OpenAI calls it a balanced model for everyday work and describes its performance as competitive with GPT-5.5. That does not mean Terra will match Sol on every difficult prompt. It means Terra is intended to offer a lower-cost balance of capability, speed, and quality across common work.

Terra is a reasonable default candidate for drafting from supplied material, workspace question answering, routine code changes, document analysis, and tool-using workflows that still involve some uncertainty. A team that starts every job on Sol may find that many prompts do not benefit enough from the extra capability. A team that sends every prompt to Luna may spend more time correcting the difficult cases. Terra sits between those two extremes.

The July 30 API price is $2 per million input tokens and $12 per million output tokens. OpenAI also says Terra remains available in ChatGPT Work, Codex, and the API. Free and Go users can access Terra in ChatGPT Work and Codex, while Plus, Pro, Business, and Enterprise users can choose Terra and Luna. Product rollouts, workspace controls, and account eligibility can still affect what an individual user sees, so the model picker in your account is the final check.

What GPT-5.6 Luna is for

Luna is positioned for speed, cost efficiency, and scale. Its July 30 API price is one tenth of Terra’s price for both input and output tokens. That gap matters when an application processes a large number of documents, runs background automations, classifies incoming material, or performs a routine step many times.

Cheap does not mean limited to a single text completion. OpenAI says Luna can use tools and complete multi-step workflows. The better boundary is task clarity. Luna is a strong candidate when the instructions, format, and acceptance test are explicit. It can handle implementation after a harder model has resolved the uncertain parts, produce structured transformations, run routine checks, or process batches where occasional escalation is built into the workflow.

A low token price can still lead to waste if prompts cause repeated retries or oversized outputs. Measure the total cost of an accepted result, including human review, failed runs, tool calls, and output length. For related workflow design, see our guide to using ChatGPT assistants for repeatable work.

Availability in ChatGPT, Codex, and the API

OpenAI made the GPT-5.6 family generally available across ChatGPT, Codex, and the OpenAI API on July 9, 2026. The newer July 30 announcement gives more specific product positioning for Terra and Luna: both remain in ChatGPT Work, Codex, and the API, with Terra available to Free and Go users and both Terra and Luna selectable on Plus, Pro, Business, and Enterprise.

The official text does not establish that every tier appears in every ordinary chat, in every workspace, or under identical controls. It also distinguishes ChatGPT Work from ChatGPT generally in the later availability details. Avoid treating an API model name, a Codex option, and a ChatGPT picker entry as interchangeable promises. If your work centers on generated documents and presentations, our ChatGPT Work guide explains that product surface. Our Codex for knowledge work overview covers a different interface and workflow.

How the July 30 pricing and credits changed

At the July 9 launch, the API prices were $5 input and $30 output for Sol, $2.50 input and $15 output for Terra, and $1 input and $6 output for Luna, all per million tokens. On July 30, Terra fell to $2 input and $12 output. Luna fell to $0.20 input and $1.20 output. Sol stayed at $5 input and $30 output.

Prompt caching has its own billing rules. OpenAI says GPT-5.6 and later models bill cache writes at 1.25 times the uncached input rate, while cache reads receive a 90 percent discount against cached input. The launch also documented explicit cache breakpoints and a minimum cache life of 30 minutes. Whether caching saves money depends on how much stable prompt material is reused.

For subscriptions, the July 30 announcement says prices and quota budgets stayed the same. Terra and Luna now use fewer credits in ChatGPT Work and Codex. OpenAI did not publish a universal conversion table in the cited announcement, so this article does not invent one.

Checklist for comparing GPT-5.6 model quality, latency, and total cost
Compare accepted results, not model names alone. Include review time, retries, latency, and token use.

A neutral model-selection framework

Begin with Luna on well-defined, repetitive work. Move to Terra when the task needs more judgment, has varied inputs, or creates too many Luna retries. Use Sol when ambiguity, long-horizon planning, difficult tool use, or the cost of failure makes additional capability worth paying for.

Test with a small set of representative cases. Keep the prompt, source material, tools, and scoring rules constant. Record factual errors, instruction failures, review time, latency, token use, and whether the output was accepted. One model can be cheaper per token but more expensive per approved result.

Routing can also happen within one workflow. OpenAI gives an example in which Sol resolves uncertainty and defines a coding plan, then Luna handles specified changes, tests, and evaluation. That pattern is useful beyond coding: use the stronger tier for the uncertain decision, then use a cheaper tier for bounded execution. Terra can replace either model when its measured results justify the middle price.

Why company benchmarks need context

OpenAI reports strong GPT-5.6 results across coding, browsing, computer use, professional work, science, and cybersecurity. These figures help describe the intended capability curve, but they are not a promise about your production workload. Some evaluations are internal. Test harnesses, graders, reasoning settings, tools, and time budgets vary. OpenAI also notes that its latency and API cost figures are simulated from production behavior and that real results may vary substantially.

Multi-agent results add another complication because total token use includes all agents while latency is derived from the root agent. A higher score may therefore involve a different cost and execution pattern. Benchmark comparisons can guide which model to test first. They should not replace a task-level evaluation using your own documents, constraints, and acceptance criteria.

Safety context and practical limits

The GPT-5.6 system card classifies Sol, Terra, and Luna as High capability for cybersecurity and biological or chemical risk, while stating that none reaches the High threshold for AI self-improvement. OpenAI says the models do not reach the Critical threshold in cyber or biological capability.

The system card describes layered safeguards, including model training, real-time checks, monitoring, and trust-based access for some sensitive defensive work. It also reports that GPT-5.6 showed a greater tendency than GPT-5.5 to go beyond user intent in agentic coding simulations, although absolute rates remained low. That is a practical reason to define permissions, confirmation points, and rollback boundaries for any model that can operate tools or edit data.

OpenAI says Sol’s cyber safeguards block roughly ten times more potentially harmful activity than previous models, which may also create friction for benign requests. Users can retry some prompts on a lower-capability model, while qualified defenders can seek Trusted Access. No safeguard makes outputs automatically correct. Human review remains necessary for consequential decisions, code deployment, health information, security work, and any action that can alter or disclose data.

Official sources

FAQ

Which GPT-5.6 model should most people start with?

For a mixed everyday workload, Terra is the clearest middle starting point because OpenAI positions it as the balanced tier. Start with Luna when the task is high volume and tightly specified. Start with Sol when the work is unusually difficult or costly to get wrong.

Is Luna available to free ChatGPT users?

The July 30 announcement says Free and Go users can access Terra in ChatGPT Work and Codex. It says Plus, Pro, Business, and Enterprise users can choose Terra and Luna. That source does not document Luna selection for Free or Go users.

Did ChatGPT subscription prices change on July 30, 2026?

No. OpenAI says ChatGPT and Codex subscription prices and quota budgets remained unchanged. Terra and Luna began consuming fewer credits, while the published API token prices for those models fell.

Does Sol always produce the best result?

No benchmark can establish that for every task. Sol is the highest-capability tier, but a clear routine prompt may not benefit enough to justify its higher cost. Compare models on representative tasks and score accepted outcomes, review effort, latency, and total cost.

OpenAI Presence Guide: How Voice and Chat Agents Work for Business

Muawia Ali
-
0
OpenAI Presence Guide: How Voice and Chat Agents Work for Business

OpenAI Presence is an enterprise product for deploying voice and chat agents in customer-facing and internal workflows. OpenAI describes it as a managed deployment that combines model reasoning with company policies, permitted actions, guardrails, evaluations, and rules for escalation to people. It is available through a limited general availability program for eligible enterprise customers.

That description matters because Presence is easy to misread as a new button or mode inside ordinary ChatGPT. It is not a normal consumer ChatGPT feature, and OpenAI says it is not yet self-serve. An organization explores access through its OpenAI account team. OpenAI Forward Deployed Engineers and selected systems integrators lead deployments.

This guide is based on OpenAI’s official Presence announcement, published July 22, 2026. It explains what the company has documented, where the announcement uses promotional language, and which commercial and technical details remain unspecified.

OpenAI Presence deployment flow from a defined business job to systems, policies, evaluation, launch, and human escalation
Presence deployments begin with a defined job and controlled access, followed by evaluation, launch, observation, and approved updates.

What OpenAI Presence is

Presence is meant to help an enterprise put an agent into a specific production workflow. OpenAI gives examples such as resolving billing issues, supporting insurance claims, and handling employee IT service requests. The agent can answer questions, use connected company systems, take actions that the organization has approved, and transfer work to a person when required.

The product supports real-time voice and chat experiences. OpenAI lists customer support, outbound sales, and higher-risk internal workflows as current contexts. A billing agent, for example, might interpret a request, verify a customer, retrieve account information, apply the company’s policy, and perform an allowed action. This example describes the intended product pattern. It does not mean every Presence agent receives every capability automatically.

Each deployment starts with one job. According to OpenAI, the agent receives only the knowledge and system access needed for that job. The customer defines what it may do, which steps need approval, and when a human should take over. This narrow scope is central to the product’s design, not an optional prompt-writing technique.

How a deployment is assembled

OpenAI says its team works with the customer to identify a useful workflow, connect relevant knowledge and systems, set permissions and policies, test the agent, and move it into production. Selected systems integrators may also support the deployment as it expands. This is a service-led implementation rather than a software download that a team configures alone.

Presence brings several components together: policies and standard operating procedures, guardrails, approved actions, simulations, evaluation tools, and a Codex-powered process for proposing improvements. Companies can keep policies, evaluation methods, and escalation rules consistent while adapting other parts for a particular workflow or channel.

The announcement does not publish a standard connector catalog or promise compatibility with named CRM, contact-center, ticketing, identity, or payment products. It says Presence can connect company systems and use account context, but the specific integration work appears to depend on the deployment. A buyer should therefore ask which systems are supported for its use case, how authentication works, what data each connection exposes, and which actions are read-only or write-enabled.

Readers comparing this managed enterprise approach with other OpenAI products may find our guide to Codex for knowledge work useful. The products can involve Codex, but they are not interchangeable: Presence is described as an agent deployment product, while the linked guide covers a different work context.

Voice, chat, and human hand-offs

Voice and chat are channels for the same basic operating model. The organization decides which behavior should remain common across channels and what should change. A spoken interaction may require different pacing or verification from a chat session, but policies and escalation conditions can be shared where appropriate.

OpenAI explicitly documents escalation to people. Guardrails can intervene when an interaction moves outside company boundaries, and tests can check whether an agent escalated at the right time. Production escalations also become a signal that teams can examine after launch. This means a hand-off is not merely a generic fallback message. It is part of the behavior that the deployment team defines and evaluates.

The source does not specify the exact transfer experience. It does not say whether every deployment supports a warm transfer, a live transcript, a generated summary, queue routing, or continuity across channels. Those details should be confirmed during solution design rather than assumed from the phrase “escalate to people.”

Testing, oversight, and changes after launch

Before release, teams can run simulations covering common requests, edge cases, and higher-risk situations. OpenAI says graders assess whether the agent reached the right outcome, followed policy, used tools correctly, and escalated when appropriate. That is a clearer standard than measuring whether an answer merely sounded fluent.

Oversight continues in production. Sessions, escalations, and quality signals can expose cases where the agent performs well or needs attention. OpenAI says Codex, using a Presence plugin, investigates those signals and suggests updates. Teams can test a proposed change against the production version before approving a staged release.

This process still requires accountable human decisions. The announcement says teams test and approve proposed changes. It does not establish that Codex autonomously edits a live agent, nor does it describe approval roles, audit-log retention, rollback mechanics, or review frequency. Enterprises should request those operational details and decide who can approve policies, permissions, tools, and production releases.

A practical evaluation should use examples from the chosen job. Tests should cover correct outcomes, prohibited actions, incomplete verification, ambiguous requests, system failures, and the point at which a person must intervene. The source supports this kind of scenario testing, but it does not provide a universal acceptance threshold. Each organization must define what adequate performance means for its own risk and service requirements.

OpenAI Presence evaluation checklist covering outcomes, policy compliance, tool use, approvals, guardrails, and human hand-offs
A useful acceptance review checks results and boundaries, including whether tools, approvals, and escalation rules worked as intended.

What OpenAI reports from early deployments

OpenAI says Presence powers its English-language phone support line at 1-888-GPT-0090. The company reports that the agent handles open-ended requests, verifies callers, uses account context, and takes approved actions. It also says the system resolves 75 percent of inbound issues without human assistance and that its improvement process reduced human hand-offs by 15 percentage points in 10 days.

Those numbers are OpenAI’s own reported results for its support channel, not an independent benchmark or a guaranteed result for another company. The announcement does not publish the underlying sample, issue mix, full grading method, or confidence intervals. A prospective customer should not convert those figures into a business case without testing its own workflow and defining comparable measures.

OpenAI also identifies BBVA, SoftBank, and IAG as enterprises exploring or testing Presence-related experiences. The wording is important. BBVA is exploring voice support for everyday banking needs in Mexico, SoftBank is testing Japanese-language conversations, and IAG is exploring support during events such as severe weather. These examples show design-partner activity; the announcement does not establish broad production rollout or identical capabilities at all three companies.

Availability and purchasing context

Presence is available to eligible enterprise customers through limited general availability. OpenAI directs interested organizations to their account teams. It says deployments are led by Forward Deployed Engineers and selected global systems integrators, with further support possible as deployment expands.

No public self-serve signup, standard price, usage rate, minimum commitment, deployment schedule, or service-level agreement appears in the announcement. It also does not state which countries, industries, languages, call volumes, or enterprise account types qualify. These are sales and scoping questions for the account team.

OpenAI separately says it will continue supporting voice customers that access frontier models through the OpenAI API. That sentence helps distinguish Presence from building directly with APIs. Presence packages deployment systems and expert involvement around a defined enterprise agent. The API remains another route for teams that are building voice products. The announcement does not provide a feature-by-feature comparison or say that one route replaces the other.

What the announcement does not establish

The official post provides a product overview, not a full contract or technical specification. It does not publish pricing, service levels, uptime commitments, data residency options, retention periods, security certifications, regulatory coverage, model choices, capacity limits, latency targets, or a complete integration list. It also does not explain whether customers can export configurations or move a deployment between providers.

None of those omissions proves that a feature or assurance is unavailable. It means the cited source does not establish it. Security, privacy, legal, procurement, and architecture teams should review current contractual and technical material for the proposed deployment instead of filling gaps with assumptions.

Presence should also not be presented as a consumer ChatGPT agent builder. If you are researching self-contained ChatGPT creation tools, our ChatGPT Sites guide covers a separate product surface. Access to ChatGPT, Codex, the API, or another OpenAI business product does not, by itself, establish eligibility for Presence.

Questions to take to an OpenAI account team

Start with the exact job, channel, users, systems, and allowed actions. Ask how caller or user verification is performed, how the agent’s permissions are limited, and how a person receives an escalated case. Request the supported integration design for each required system rather than relying on a general claim that systems can be connected.

Then ask how simulations and graders are created, who can change policies, how releases are approved, and what evidence is retained. Commercial discussions should cover eligibility, implementation responsibilities, partner involvement, pricing, support, service commitments, data handling, and exit arrangements. These questions do not imply answers that OpenAI has not published. They turn the public overview into a concrete due-diligence process.

Frequently asked questions

Is OpenAI Presence a feature in regular ChatGPT?

No. OpenAI describes Presence as a deployed enterprise product in limited general availability. It is not self-serve, and interested organizations are told to contact their OpenAI account team.

Does Presence support both voice and chat agents?

Yes. OpenAI says the product currently supports real-time voice and chat experiences. The customer decides which policies, evaluations, and escalation rules stay consistent and what changes for each workflow or channel.

Can a Presence agent take actions in company systems?

OpenAI says an agent can use connected systems and take approved actions. Access is scoped to the job, while the company defines permissions, approval points, and hand-off rules. The announcement does not list every supported integration or action.

How can an organization get OpenAI Presence?

Eligible enterprise customers can access it through a limited general availability program. OpenAI says to contact the organization’s account team. The public announcement does not provide self-serve enrollment, standard pricing, or universal eligibility criteria.

Bottom line

Presence is best understood as a managed enterprise deployment for a narrowly defined voice or chat job. Its documented model combines limited knowledge and system access with company policies, testing, guardrails, approved actions, human escalation, and reviewed updates after launch. The official announcement establishes that operating model and the current access route. Detailed commercial, security, integration, and service commitments must come from the materials supplied for a specific deployment.

Finances in ChatGPT: Connecting Accounts, Privacy, and Practical Limits

Muawia Ali
-
0
Finances in ChatGPT: Connecting Accounts, Privacy, and Practical Limits

Finances in ChatGPT is a dedicated personal finance experience that can connect supported financial accounts, organize account data into a dashboard, and use that context when answering finance questions. It is not a system for moving money or making trades. It also does not turn ChatGPT into a financial professional.

Financial disclaimer: ChatGPT is not a fiduciary, registered investment adviser, broker-dealer, tax preparer, law firm, or substitute for a qualified professional. Its responses are for information and planning only and may contain mistakes. Check account data and calculations yourself. Consult an appropriately qualified professional before making a decision with legal, tax, credit, retirement, or investment consequences.

This guide follows OpenAI’s current documentation rather than relying on a third-party report. It explains who can use the feature, what happens during a Plaid connection, what ChatGPT can and cannot do, and how to remove connected data. Availability and behavior may change, so check the linked OpenAI pages before relying on any detail.

Who can use Finances in ChatGPT?

OpenAI’s current Help Center article says Finances is available to ChatGPT Plus and Pro users in the United States on the web, iOS, and Android. The rollout is gradual, so an eligible user may not see it immediately. The feature is not documented there as available to Free users, users outside the United States, or other plans and workspace types.

The distinction matters because the original May 15, 2026 announcement described a preview for Pro users in the United States, initially on web and iOS. OpenAI added an update dated June 25, 2026 saying the experience had expanded to Plus and Pro users in the United States on web, iOS, and Android. The Help Center reflects that newer eligibility. If the Finances entry is missing from your sidebar despite meeting those conditions, gradual rollout may be the reason.

OpenAI calls this experience a new personal finance experience in ChatGPT. Its separate Finances in ChatGPT Help Center guide contains the more detailed operating, privacy, removal, and troubleshooting information used below.

What the feature can show

Once supported accounts have synced, the Finances page provides a dedicated dashboard. What appears depends on the data your institutions make available and whether ChatGPT has enough information to generate a particular widget. OpenAI lists possible views for spending by category, current-month spending, upcoming payments, subscriptions, bills, net worth, portfolio distribution, and market updates for equity holdings.

The dashboard also shows connection status and the last sync time. A pencil icon lets you reorder widgets or choose which widgets appear. A missing card does not necessarily indicate an error. It may mean the connected institution did not provide the needed fields or that ChatGPT lacks enough data.

Finances in ChatGPT account connection and dashboard workflow

In chat, connected context can support questions about where money went, recurring charges, expected bills, spending trends, net worth, and portfolio allocation. ChatGPT may cite the relevant connected data so you can inspect the basis for an answer. You can ask questions from the Finances page or in a regular conversation by using /finances or @finances.

How the Plaid connection works

Open Finances from the ChatGPT sidebar and select Get started. You can also type @Finances, connect my accounts in a conversation. ChatGPT then offers Connect with Plaid. After that, sign in through the connection flow, choose the financial institution and accounts you want to connect, and return to ChatGPT.

Plaid maintains connections to more than 12,000 financial institutions, including supported banks, credit card issuers, brokerages, and other providers. That number does not mean every institution or every account type will work with Finances. An institution may be absent, temporarily unavailable, or unable to supply the particular data the feature needs.

After authentication, ChatGPT starts syncing and categorizing data. OpenAI says this usually takes a few minutes, although rare cases can take several hours. Syncing continues automatically. If an account needs attention, the Accounts tab may ask you to reconnect, update login information, or complete a security step with the institution.

Choose only the accounts and data you are comfortable sharing. According to OpenAI, ChatGPT can access balances, transactions, investments, and liabilities when those data are supplied through the connection. It cannot see full account numbers. The feature also cannot change an account, move money, or place an order.

Permissions and practical limits

Finances is an information and planning surface, not a financial transaction service. OpenAI says ChatGPT cannot move money, pay bills, change account settings, make trades, change retirement contributions, open or close financial accounts, or file taxes. It cannot act as your financial, legal, tax, or investment adviser.

Connected data can also be incomplete. One institution may provide balances without transaction history. Another may omit holdings, loan details, annual percentage rates, due dates, or other fields. Historical coverage is limited to what the institution supplies. ChatGPT cannot reconstruct missing records merely because an account connected successfully.

Classification introduces another source of error. Transfers, reimbursements, credit card payments, and duplicate pending transactions can be counted as spending or placed in the wrong category. Dashboard figures can also lag while an account is syncing or when an institution has not shared its latest data. Ask which transactions and fields were used, then compare the answer with the institution’s own records.

You can ask ChatGPT to reclassify a transaction or explain why a widget differs from your expectation. Treat that as a correction to the conversational context, not a change to the bank’s records. For unsupported property, private loans, planned purchases, or other manual details, you can describe the item in a finance conversation and decide whether to save it as a financial memory.

Financial memories and shared context

Account feeds do not capture every relevant fact. You might share that rent is divided with a partner, that a family loan remains outstanding, or that you want to keep a minimum balance in checking. ChatGPT can save such details as financial memories and use them in later finance conversations.

Financial memories are separate context intended for finance conversations. You can view or delete them from the Finances page through the three-dot menu. If you want a broader explanation of how saved context works, see our guide to ChatGPT memory and its controls. Remember that deleting a connected account and deleting a financial memory are different actions.

Privacy, training controls, and deleting data

Existing ChatGPT data controls apply to Finances. Conversations involving connected finance data follow the model training setting selected for the account. You can review or change that choice under Settings > Data controls. OpenAI does not describe financial conversations as automatically exempt from the training preference you have chosen.

Temporary Chat behaves differently. In a temporary chat, ChatGPT will not access connected financial accounts, use memories for personalization, or create memories. The temporary conversation also does not appear in chat history. This can be useful when you want a finance-related discussion that does not draw on the connected account context.

To disconnect Finances, use Settings > Apps > Finances or the Finances home page. The Accounts tab also lets you manage or remove individual connections. OpenAI says synced account data is deleted from its systems within 30 days after disconnection. Data associated with that ChatGPT connection is deleted from Plaid within 30 days under Plaid’s policies.

Disconnection does not remove financial information already present in your ChatGPT conversation history. Delete those conversations separately if you no longer want them retained in history. It also does not remove financial memories, which have their own controls. Finally, disconnecting Finances affects the ChatGPT connection only. It does not cancel other Plaid connections you may have authorized for unrelated apps.

Checklist for reviewing Finances in ChatGPT data, memories, and privacy controls

Useful questions to ask

Start with questions whose inputs you can verify. Ask, “Which transactions did you count as dining this month?” before asking for a spending comparison. For recurring costs, try, “List the subscriptions you found, including amount, cadence, and next expected charge date.” A bill review can begin with, “Which upcoming payments are based on confirmed recurring transactions, and which are estimates?”

For a cash-flow check, ask ChatGPT to separate account data, memories, and assumptions. You might request a comparison of this month’s grocery spending with recent months, then inspect any transfers or reimbursements included in the total. For debt, ask it to organize balances and interest rates, identify missing fields, and model several payment amounts without presenting one scenario as a recommendation.

Portfolio questions need similar care. Ask for the allocation shown by the available holdings and the date of the last sync. Do not assume the result includes every brokerage, private asset, retirement account, or current market price. A daily change shown in a widget is not a reason by itself to trade.

For tax topics, ChatGPT can organize information, explain concepts, and point out areas to review, such as estimated payments or possible deductions. OpenAI states that it does not file taxes and does not replace a tax professional. Do not submit a return or execute a tax-sensitive transaction solely from a generated response.

A careful review routine

  1. Check the last sync time and resolve any account marked as needing attention.
  2. Confirm which accounts are connected and whether an expected account is missing.
  3. Ask ChatGPT to identify the source data and assumptions behind important figures.
  4. Compare balances, transactions, holdings, dates, and calculations with original statements.
  5. Review financial memories and remove details that are outdated or no longer useful.
  6. Before disconnecting, delete any conversations and memories you also want removed.

Apply the same caution you would use with any sensitive ChatGPT connection. Our ChatGPT Health privacy-controls guide covers a different product, but its practical habit is relevant here: inspect permissions, understand what persists, and verify consequential information at the original source.

FAQ

Is Finances in ChatGPT available to every user?

No. OpenAI currently documents availability for Plus and Pro users in the United States on web, iOS, and Android. Rollout is gradual, so some eligible accounts may not see it immediately.

Can ChatGPT transfer money or trade from a connected account?

No. OpenAI says it cannot move money, pay bills, make trades, change retirement contributions, alter account settings, or open and close accounts. The connection supplies context for viewing, explanation, and planning.

What happens after I disconnect Finances?

OpenAI says synced account data is deleted from OpenAI’s systems within 30 days, and data associated with the ChatGPT connection is deleted from Plaid within 30 days under Plaid’s policies. Conversation history and financial memories require separate deletion.

Should I rely on a Finances answer for an investment or tax decision?

No. Review the underlying data and consult a qualified professional when a decision has investment, tax, legal, credit, or retirement consequences. ChatGPT may make mistakes and does not act as an adviser, tax preparer, fiduciary, broker-dealer, or law firm.

Bottom line

Finances in ChatGPT can place supported account data and personal context in one conversational workspace. Its value depends on the completeness and freshness of the connected data, careful memory management, and the user’s willingness to check every consequential answer. Connect selectively, inspect the source records, and keep financial decisions in human hands.

CHATGPT Custom Instructions Guide: How to Use the New 5,000-Character Limit

Muawia Ali
-
0
CHATGPT Custom Instructions Guide: How to Use the New 5,000-Character Limit

Custom instructions give ChatGPT a standing set of preferences to consider when it responds. They are useful for choices that should remain consistent across conversations, such as your preferred tone, reading level, units, formatting habits, or the kind of context you regularly need. On July 15, 2026, OpenAI increased the available space for some paid and organizational plans from 1,500 to 5,000 characters. The larger field makes room for clearer guidance, but it does not turn custom instructions into a private database or a guarantee that every response will follow every line.

This guide explains who receives the larger limit, what custom instructions affect, how to write them well, and what to consider before storing ongoing context. It is based on OpenAI’s current help documentation and release notes, not on personal product testing.

Who can use the 5,000-character limit?

OpenAI’s Custom Instructions help article lists two limits. Plus, Pro, Enterprise, Business, and Education users can save up to 5,000 characters. Free and Go users can save up to 1,500 characters. Custom instructions themselves are available on all plans across web, desktop, iOS, and Android; the difference is how much text each eligible account can store.

The July 15, 2026 release note describes the increase as a change from 1,500 to 5,000 characters for the five eligible plan groups. It says the added room is intended to support more customization of response style and behavior. It does not announce 5,000 characters for Free or Go accounts, so users on those plans should design within the documented 1,500-character allowance.

A character limit counts more than visible words. Spaces and punctuation also occupy space. You do not need to fill the field, and a shorter set of specific instructions is often easier to maintain than a long policy document.

What custom instructions change

OpenAI describes custom instructions as information you want ChatGPT to consider in its responses. That can cover stable preferences and recurring context. For example, you might ask for plain English, metric units, concise answers first, code examples in a particular language, or a short explanation of unfamiliar terms.

These instructions act as background guidance. They do not replace the request you type in a chat. A useful division is to keep durable preferences in custom instructions and put task-specific details in the current prompt. If you are writing one grant application, the grant rules belong in that conversation. If you always want dates written in ISO format, that may belong in custom instructions.

OpenAI says updates are applied immediately across chats, including existing conversations. In practice, that means subsequent responses can use the new version. Earlier messages are not rewritten. The help article also notes that previous versions may remain visible in chat history. Removing or editing the setting controls future responses, while clearing an old conversation is the documented way to remove old instruction text that appears in that conversation.

Workflow for separating permanent ChatGPT custom instructions from task-specific prompts

How to find and control the setting

Use the path documented for your platform rather than relying on an old screenshot. On web and desktop, OpenAI directs users to Settings, then Personalization. Make sure Enable customization is on and enter the text in the Custom Instructions field. On iOS and Android, open Settings, select Customize ChatGPT, turn on Enable customization, and enter the instructions in the same named field.

You can disable the feature without immediately deleting the saved text. On mobile, turn Enable customization off under Customize ChatGPT. On web and desktop, go through Personalization and Custom Instructions, then turn the setting off. OpenAI presents deletion as an optional separate step. If you only want to compare behavior temporarily, disabling first is less destructive than erasing a carefully edited set.

Design instructions around stable decisions

The extra space works best when it removes repeated setup rather than adding a biography. Start by listing decisions you make in many chats. Then keep only the ones that would still be useful a month from now.

A practical structure is:

  • Context: your role or audience, limited to details that materially change an answer.
  • Response preferences: tone, depth, units, language, accessibility needs, or preferred output format.
  • Working rules: how to handle uncertainty, assumptions, citations, calculations, or code.
  • Boundaries: habits you do not want, such as invented sources, unnecessary jargon, or a conclusion that repeats the introduction.

Write each line so that another person could tell whether it was followed. “Write well” gives little direction. “Use short paragraphs, define technical terms on first use, and put the answer before background” is easier to interpret. Avoid packing several unrelated requirements into one sentence. If two rules conflict, revise them instead of hoping the model will decide which one matters more.

A reusable example for everyday work

The following is an original example, not an official OpenAI template:

I work with a mixed technical and nontechnical audience. Start with a direct answer, then add the reasoning needed to act on it. Use plain English and short paragraphs. Define specialized terms the first time they appear. Use metric units and ISO dates. When information is uncertain, state the uncertainty and identify what would confirm it. Do not invent quotations, sources, test results, or product behavior. For recommendations, explain the main tradeoff instead of presenting one choice as universally best. Use tables only when they make comparison easier.

This example is compact because each sentence addresses a recurring choice. It does not include a current project name, client data, a temporary deadline, or the contents of a document. Those belong in the relevant conversation, where they can be removed from the standing context later.

Use the larger field without creating contradictions

Five thousand characters can support more detail, but length is not a quality score. Long instructions often fail for ordinary editorial reasons: the same idea appears twice, exceptions are scattered, or a new rule conflicts with an old one.

Group related requirements and put the most important ones first. Prefer a clear default with one necessary exception. For example, “Answer concisely by default; give a detailed derivation when I ask for one” is more coherent than separate commands demanding both maximum detail and maximum brevity.

Keep examples limited to cases that clarify a rule. Ten formatting samples can consume space while making the intended default less obvious. If a project needs extensive terminology, a long rubric, or many source documents, keep that material with the project or conversation instead of turning account-wide instructions into a storage area.

Privacy, data use, and sensitive permanent context

Custom instructions persist as account-level settings, so treat them as ongoing context rather than a disposable prompt. Do not store passwords, API keys, payment details, private medical records, confidential client material, or other information that does not need to be present in every chat.

OpenAI says custom instructions are not shared with viewers of a shared link. It also warns that when third-party plug-ins are used, the model may give plug-in developers relevant information from the instructions. OpenAI therefore advises using only plug-ins you trust and excluding information you would not want provided to those developers.

The help article says information from the use of custom instructions may be used to improve model performance and links to OpenAI’s training opt-out information. Review your account’s Data Controls if you do not want your content used for that purpose. The same help page confirms that custom instructions are included in a ChatGPT data export.

Deleting the setting and deleting the account are different actions. You can edit or delete instructions for future conversations at any time. If you delete your OpenAI account, OpenAI says associated custom-instruction data is deleted within 30 days as part of that process. It also notes that chat records may have been de-identified and disassociated from the account if you had not opted out of content use.

If you want a broader look at related personalization controls, read our guide to ChatGPT memory and its controls. For account-level privacy choices, our ChatGPT training opt-out guide covers that separate setting.

Checklist for reviewing ChatGPT custom instructions, privacy, and conflicting rules

What custom instructions cannot guarantee

Custom instructions shape responses, but they do not verify facts, create new product permissions, or make unsafe information appropriate to share. A line asking for citations does not prove that a citation is real. A request for accurate calculations does not remove the need to check consequential results. Review important output against reliable sources.

The feature is also separate from the API. OpenAI’s help article says there is no custom-instructions API and directs API developers to use system messages for a similar effect in the Chat Completions API. Do not assume that text saved in the ChatGPT interface automatically becomes configuration for an API application.

Custom instructions can also collide with a direct prompt or with each other. When the output becomes awkward, inspect the standing rules before blaming the current question. A requirement to avoid lists may conflict with a request for a checklist. A general preference for brief answers may be a poor fit for a detailed contract comparison. The current task still needs enough context to resolve exceptions.

How to revise your instructions

Review the setting when your role, audience, tools, or writing needs change. A simple revision process keeps the text useful:

  1. Save a private copy of the current text if you may want to restore it.
  2. Remove outdated project names, repeated rules, and context that no longer affects most chats.
  3. Turn vague preferences into observable instructions.
  4. Resolve pairs that ask for opposite behavior.
  5. Disable customization briefly if you need to compare responses without the standing guidance.
  6. After editing, use future responses to see whether the important defaults are clear. Revise the instruction itself rather than adding more exceptions to every prompt.

Do not treat one imperfect answer as proof that the entire setup needs another thousand characters. First check whether the relevant rule was specific, whether the current prompt supplied the needed facts, and whether another standing instruction competed with it.

Frequently asked questions

Does every ChatGPT plan get 5,000 characters?

No. OpenAI currently documents 5,000 characters for Plus, Pro, Enterprise, Business, and Education. Free and Go users can save up to 1,500 characters.

Do changes affect an existing conversation?

OpenAI says setting updates apply immediately across all chats, including existing conversations. That affects subsequent responses. It does not rewrite earlier messages, and previous instruction text may still appear in chat history.

Are custom instructions private when I share a chat link?

OpenAI says they are not shared with shared-link viewers. However, relevant information may be provided to third-party plug-in developers when you use their plug-ins, so avoid sensitive standing context and use only plug-ins you trust.

Should I use all 5,000 characters?

No. Use enough space to state stable, non-conflicting preferences. Keep temporary project details in the current conversation and remove rules that no longer improve typical responses.

A short final checklist

Confirm that your account has the limit you expect. Keep standing preferences separate from current-task details. Write rules that can be checked, remove contradictions, and avoid sensitive information that does not belong in every chat. Review Data Controls and plug-in access before saving account-wide context. Most importantly, revise the instructions when they stop matching the way you actually work.

ChatGPT Memory and Controls: A Practical Guide for 2026

Muawia Ali
-
0
ChatGPT Memory and Controls: A Practical Guide for 2026

ChatGPT Memory is no longer just a small list of facts that you explicitly ask the assistant to save. OpenAI’s current system can build a changing summary from useful context across chats, files, and connected apps, then use that context to personalize later answers. The same feature can be convenient and easy to misunderstand. Turning memory off, deleting a chat, clearing the memory summary, and disabling model training are separate actions with different results.

This guide explains the controls OpenAI documents in July 2026. It is based on OpenAI’s help pages and release notes, not personal account testing. The interface can vary while updates roll out, so use the labels available in your own Settings page and avoid assuming that another user’s screen will match yours.

What ChatGPT Memory means in 2026

When memory is enabled, ChatGPT may use information from your chats, uploaded files, and connected apps to make future replies more relevant. OpenAI describes the current system as a continually updated synthesis rather than a complete transcript or a fixed notebook. You can review a memory summary, correct it, and delete the memories shown there. The summary can say when it was last updated, but it may not display every detail that affects personalization.

This current experience differs from the legacy saved memories system. Saved memories are individual details that you directly ask ChatGPT to remember, or that ChatGPT saves because they may help in later conversations. They are stored separately from chat history and remain part of future context until removed. The newer memory summary updates more automatically and is meant to avoid the stale or contradictory items that can accumulate in a manual list.

OpenAI still documents a way to return to legacy saved memories through Settings > Memory > Saved memories. The exact controls you see depend on whether the updated experience has reached your account. According to the ChatGPT release notes, memory summary updates were still rolling out across plans and countries in June 2026. That is the safest way to describe availability: documented rollout, not universal visual parity.

Diagram comparing ChatGPT memory summary, saved memories, and past chat sources
Memory can draw on several sources, while each source still has its own deletion and privacy controls.

Memory summary, saved memories, and past chats are different

The memory summary is a high-level view of what ChatGPT currently considers useful about you. It changes as you keep chatting. OpenAI says the summary should contain the most important details, but it is not a complete ledger. If you want to check whether a particular detail has been remembered, OpenAI suggests asking ChatGPT directly.

Legacy saved memories behave more like persistent notes. You might say, “Remember that I prefer vegetarian recipes,” and ChatGPT can consider that preference later. OpenAI says ChatGPT may also save useful details without an explicit request. You can delete an item, clear all saved memories, or turn the feature off. Deleting the chat where a saved memory began does not necessarily delete that memory because the two are stored separately.

Past chat context is broader. In the updated experience, ChatGPT can synthesize useful information from conversation history rather than relying only on a short list of saved items. This does not mean every past message is inserted into every new prompt, nor does the memory summary show every factor. It means relevant context from previous conversations may influence a response when memory is enabled.

Chat history itself remains a separate account record. You can search and reopen old conversations using the current interface described in our ChatGPT unified search guide. Finding a chat, deleting a chat, and changing memory are three different operations.

How to review and correct what ChatGPT remembers

Open Settings, choose Personalization, then open Memory. OpenAI’s Memory FAQ documents the memory summary and its management options. You can type a requested change into the box below the summary, or highlight text in the summary and make a specific correction. If the summary is empty, the account may be new, memory may have been enabled recently, or there may not yet be enough history. A Refresh option is available from the summary’s menu.

Corrections matter because personalization can be based on context that has changed. A location, work role, diet, current project, or preference may no longer be accurate. State the correction plainly and then review the summary. Do not treat a fluent response as proof that every stored detail has been changed.

OpenAI also provides source visibility beneath personalized responses. Tap the book icon below a response to see sources such as custom instructions, past chats, files, and memories. Selecting a memory can explain why it was used, and the three-dot menu can offer a correction route. OpenAI cautions that this source view may not show every factor that shaped the answer. Memory sources are not included when you share a chat.

Deleting a memory is not the same as deleting a chat

If you delete a chat, a legacy saved memory created from that conversation may still remain. If you delete a saved memory, the old chat can still contain the original statement and earlier replies that mention it. OpenAI says deleted saved memories may be retained in a log for up to 30 days for safety and debugging. Removing a memory also does not rewrite text that already exists in your conversation history.

The current memory summary adds another layer. The “Delete and turn off memory” option deletes the memories shown on the summary page and disables memory, but it does not delete past chats. If you enable memory again later, ChatGPT may form new memories from conversations that still exist, including older chats. That control was available on the web and rolling out on mobile when OpenAI documented it in June 2026.

For a more complete removal, delete every source where the information appears. OpenAI lists past chats, archived chats, files, the memory summary, and connected apps that contain the detail. Disconnecting an app prevents an ongoing connection, while deleting relevant stored material addresses information already present in your account. Check all of these surfaces rather than relying on one toggle.

Turning off training does not turn off Memory

Data Controls govern whether conversations help improve OpenAI’s models. Memory governs personalization. They are related to your data, but they are not the same setting. Under Settings > Data Controls, turning off “Improve the model for everyone” keeps regular conversations in your history while preventing them from being used to train ChatGPT. OpenAI says this preference applies across devices for the account.

Turning off model training does not automatically clear memory or stop past context from personalizing replies. Turning off memory does not necessarily delete chats or change the model-training preference. Review both sections when your goal involves both personalization and training. OpenAI’s Data Controls FAQ is the current source for the training toggle, account export, and related controls.

Custom Instructions are separate again. They are explicit directions about what ChatGPT should know or how it should answer. OpenAI says Temporary Chat still follows enabled custom instructions, even though it does not use or create memories. Review custom instructions if you want a conversation with neither remembered personal context nor standing instructions.

Checklist for reviewing ChatGPT memory, chat history, training, and Temporary Chat controls
A privacy review should check each control separately because one switch does not clear every source.

When to use Temporary Chat

Temporary Chat is the clearest option when you do not want a conversation to use existing memories or create new ones. Start a new chat and select the pill-shaped Temporary button in the upper-right corner, as documented in OpenAI’s Temporary Chat FAQ. Temporary chats do not appear in chat history and are not used to improve OpenAI’s models.

Temporary does not mean that no copy can exist. OpenAI may retain a copy for up to 30 days for safety purposes, and limited safety-relevant context may still be used in rare, high-risk situations. If you use a custom GPT with actions, information sent to an outside service is governed by that recipient’s privacy policy and may be kept longer. Temporary Chat cannot override a third party’s data handling.

Use Temporary Chat for a one-off discussion that should not personalize later answers. It is not a reason to paste secrets, medical records, financial credentials, workplace data, or another person’s private information without authorization. The safer rule is to minimize sensitive details before sending them at all.

Sensitive information and practical privacy limits

OpenAI says sensitive information may appear in memory if you share it with ChatGPT. The product controls reduce unwanted reuse, but they do not convert unnecessary disclosure into a good practice. Remove names, account numbers, addresses, passwords, access tokens, health identifiers, and confidential business details when they are not needed for the task.

Connected apps and uploaded files deserve the same attention as typed messages. A memory may be influenced by a file or app even if the detail is not obvious in the visible chat. The source icon can improve transparency, but OpenAI says it may not show every source or factor. If something must be fully removed, inspect the original file, app connection, archived conversations, and memory summary.

For project-based work, keep persistent reference material in a clearly scoped location rather than expecting general memory to act as a document archive. Our guide to ChatGPT project sources explains that separate workflow. Memory is for personalization context, not a guaranteed record of every fact or file.

A short control review

  1. Open Settings > Personalization > Memory and review the summary available to your account.
  2. Correct details that are outdated or wrong, then ask whether the specific fact is still remembered.
  3. Use the book icon under a personalized response to inspect the sources OpenAI exposes.
  4. Delete both the memory and the original chats, files, or app sources when complete removal is the goal.
  5. Review Data Controls separately and set “Improve the model for everyone” according to your preference.
  6. Use Temporary Chat when a conversation should neither access nor create personalization memories.
  7. Check custom instructions too, since Temporary Chat can still follow them.

Frequently asked questions

Does deleting a chat delete the memory it created?

Not necessarily. Saved memories are stored separately from chat history. Delete the saved memory or summary entry as well as the chat, and remove any other file, archived chat, or connected-app source that contains the same information.

Can I see everything ChatGPT remembers about me?

The memory summary gives a high-level view, but OpenAI says it may not contain every detail or factor used for personalization. You can ask ChatGPT whether it remembers a specific fact and inspect the book icon under a personalized response.

Does turning off “Improve the model for everyone” disable Memory?

No. The Data Controls setting governs model training, while Memory governs personalization. A regular chat can remain in history and still participate in personalization even when it is not used for training, depending on your memory settings.

Is Temporary Chat completely unrecorded?

No. It does not appear in history, use or create personalization memories, or train OpenAI’s models, but OpenAI may keep a copy for up to 30 days for safety. Data sent through a GPT action may follow a third party’s separate retention policy.

What to remember

ChatGPT now combines a current memory summary with context from past conversations, while the legacy saved memories view remains available for accounts that use it. The controls are useful only when you distinguish their jobs. Memory changes personalization. Deleting chats changes history. Data Controls change training preferences. Temporary Chat avoids personalization memory for one conversation, subject to documented safety retention and third-party limits.

123...13Page 1 of 13