OpenAI Release Notes

Codex and ChatGPT feature updates

• Choose a new URL for a ChatGPT Site.
  Site owners on Plus and Pro can change an existing Site’s ChatGPT-hosted URL without redeploying it. The previous address redirects to the new URL, including routes and query parameters, while custom domains remain separate and unchanged. In the Site’s settings, select Change URL.

• Use Apple Messages from Codex and ChatGPT Work.
  On Apple silicon Macs, the Apple Messages plugin in the ChatGPT desktop app can read and search iMessage, SMS, and RCS conversations and prepare or send messages through Messages. By default, ChatGPT asks the user to approve the message and recipients before sending.

• Use Computer History in more regions.
  Computer History is now available in the EEA, Switzerland, and the United Kingdom for Pro users in the ChatGPT desktop app on macOS. It is off by default and requires Memories.

• Share a read-only snapshot of a Codex chat.
  Personal share links can be opened by anyone with the link. A snapshot is static and does not include tool calls or shell input and output. Codex redacts known secret patterns, but users should review the snapshot because sensitive paths, diffs, images, or other content may remain.

• Keep pinned chats in sync.
  Pinned chats now stay in sync between the ChatGPT desktop app and iOS when you use the same Codex account. Android is not included in this release.

Enterprise feature updates

• Workspace owners and admins can use new workspace-scoped Admin APIs to automate invitation and member administration. Invitation APIs support listing, retrieving, creating, resending, and deleting invitations. User APIs support retrieving accepted members and, for workspace owners, updating built-in roles or seats and removing workspace memberships. SCIM-managed changes continue through the identity provider, and last-owner protections apply. For setup and permissions, see Managing Admin keys. For endpoint paths and schemas, see the Admin API reference.

• The Global Admin Console now makes it clearer which tenant or workspace you are working in and adds workspace-specific URLs for supported pages. Workspace members can view analytics limited to their own activity, while cross-user views remain available only to authorized roles. ChatGPT Workspace Settings remain in their current location. Learn more about the Global Admin Console.

• Personal Analytics supports Work and Codex for eligible Enterprise and Edu workspaces. After an admin enables the plugin, members can install it in ChatGPT Desktop and ask about activity associated with their signed-in identity and privacy-protected aggregate workspace trends. It is read-only and cannot show another user’s individual activity or another workspace’s data. Learn more about Personal Analytics.

• Collaborate on ChatGPT Sites (Enterprise only). In Enterprise workspaces, Site owners can add active members of the same workspace as editors. Editors can update and save a Site and, after the owner publishes it for the first time, publish later versions to the same Site URL. Owners keep control of sharing, Site settings, analytics, ownership, version restoration, and editor access. Co-editing uses the existing Sites workspace controls and does not add a separate RBAC toggle; public publishing remains off by default and admin-controlled. Learn more about managing ChatGPT Sites for your workspace.

• Use Apple Messages from Codex and ChatGPT Work. On Apple silicon Macs, the Apple Messages plugin in the ChatGPT desktop app can read and search iMessage, SMS, and RCS conversations and prepare or send messages through Messages. By default, ChatGPT asks the user to approve the message and recipients before sending. Managed-workspace admins can disable Apple Messages through the existing Computer Use control. Learn more about the Apple Messages plugin.

• Use Computer History in more regions (Enterprise only). Computer History is now available in the EEA, Switzerland, and the United Kingdom for Enterprise users in the ChatGPT desktop app on macOS. It is off by default, requires Memories, and must be enabled by a workspace admin before a member can opt in. Learn more about Computer History.

• Share a read-only snapshot of a Codex chat. Workspace share links are limited to authenticated members of the originating workspace, and admins can turn off workspace share links. A snapshot does not update when the original chat changes. Codex redacts known secret patterns, but users should review the snapshot because sensitive paths, diffs, images, or other content may remain. Learn more about sharing a Codex chat.

• Export the public plugin catalog. Eligible ChatGPT Enterprise workspace owners and admins can export a CSV from Admin > Plugins > Public > Export CSV. The export can be up to 48 hours old, and excludes workspace-created plugins. Learn more about exporting the public catalog.

New Features

• Added an interactive codex agents dashboard for searching, starting, opening, renaming, and stopping tasks, with configurable shortcuts. (#39094, #39112, #39114, #39142)
• Added /cd, /pwd, and /cwd commands for managing the working directory in TUI sessions. (#38894)
• Added codex queue for sending messages to existing local or remote sessions. (#39092)
• Expanded Vim editing with character replacement and more change motions such as cw, c$, and cc. (#39661)
• codex doctor now diagnoses endpoint protection, network/proxy failures, desktop app state, and update connectivity. (#38827, #38918, #39060, #39074)
• SDK users can now pass exact CLI config overrides and select max or ultra reasoning effort. (#38817, #39662)

Bug Fixes

• Queued messages now wake idle sessions reliably, resolve duplicate session names more usefully, and preserve pasted or deferred command semantics. (#39034, #39385, #39604)
• Resumed and forked threads now restore their active permission profile instead of silently falling back to current defaults. (#39153)
• Fixed duplicate sub-agent activity and tightened TUI routing for sub-agent notifications and approvals. (#39049, #39088)
• Realtime WebRTC sideband connections now reconnect after unexpected transport loss without dropping pending output. (#39257)
• Inline TUI history now remains available in Windows Terminal scrollback. (#39619)
• Inactive TUI thread replay buffers are now bounded to prevent excessive retained streamed output. (#39081)

Documentation

• Clarified that external contributions should go through issues and design discussion rather than pull requests. (#39089)
• Documented DNS exfiltration risks and trust limitations for secure devcontainers. (#39283)

Changelog

Full Changelog: rust-v0.148.0...rust-v0.149.0

• Add raw config overrides to the TypeScript SDK (#38817) @copyberry
• Support metadata staging for reserved thread IDs (#38819) @copyberry
• Avoid cloning TUI history span content (#38822) @copyberry
• Avoid allocating per character when decorating hyperlinks (#38823) @copyberry
• Add endpoint protection checks to codex doctor (#38827, #38918, #39060, #39074) @copyberry
• Isolate external editor buffers from sandbox-writable paths (#38830) @copyberry
• Share editor keymaps across TUI composer components (#38837) @copyberry
• Identify Mac mini hosts in remote control handshakes (#38840) @copyberry
• Restore thread timestamp maxima independently (#38893) @copyberry
• Add working-directory commands to the TUI (#38894) @copyberry
• Move requirements policy ownership to execpolicy (#38899) @copyberry
• Honor per-environment shell variable policies (#38902) @copyberry
• Edit queued messages with Vim history-up (#38907) @copyberry
• Stop rendering columns after filling their area (#38913) @copyberry
• Honor legacy :project_roots permission entries (#38916) @copyberry
• Improve codex doctor network diagnostics (#38918) @copyberry
• Reject obsolete app-server permission profile fields (#38919) @copyberry
• Compact successful command activity in the TUI (#38921) @copyberry
• Add experimental app-server project APIs (#38940) @copyberry
• Prevent Noise auth tokens from reaching child processes (#38941) @copyberry
• Enforce environment-specific command policies (#38942) @copyberry
• Stop loading legacy managed config on Windows (#38947) @copyberry
• Add a configurable skill catalog token budget (#38978) @copyberry
• Bound Guardian v2 parent compaction context (#38980) @copyberry
• Include images in Guardian v2 transcripts (#38987) @copyberry
• Source Guardian v2 defaults from the model catalog (#38990) @copyberry
• Evaluate recent and character-routed skill selection (#38993) @copyberry
• Remove the workspace settings gate for apps and plugins (#38994) @copyberry
• Retry transient Guardian sampling failures (#38997) @copyberry
• Ignore stale Guardian tool risk scores (#39001) @copyberry
• Disable Guardian V2 for managed automatic reviewers (#39005) @copyberry
• Emit API-key turn cost telemetry from app-server (#39006) @copyberry
• Wait for Guardian score progress in approval tests (#39007) @copyberry
• Add task context to shadow skill selection (#39008) @copyberry
• Move MCP policy types into codex-protocol (#39017) @copyberry
• Isolate MCP resource headers during OAuth requests (#39019) @copyberry
• Restrict Windows IDE pipe client impersonation (#39020) @copyberry
• Hang soft-break whitespace in the TUI composer (#39022) @copyberry
• Reduce retained MCP result data in TUI history (#39023) @copyberry
• Speed up the large diff highlighting test (#39025) @copyberry
• Use simulated time in the long typing test (#39026) @copyberry
• Cache the test pet spritesheet encoding (#39028) @copyberry
• Speed up the in-app resume cwd test (#39030) @copyberry
• Encapsulate TUI shimmer under the motion module (#39032) @copyberry
• Bound legacy resume preview scans (#39033) @copyberry
• Dispatch queued messages written by other processes (#39034) @copyberry
• Add app-server coverage for Guardian V2 approval routing (#39035) @copyberry
• Allow config reads to join active app-server read batches (#39036) @copyberry
• Restore Guardian risk scores across thread lifecycles (#39038) @copyberry
• Preserve foreign paths in managed network approvals (#39040) @copyberry
• Enforce managed authentication backend settings (#39043) @copyberry
• Add managed gates for in-app chat and dictation (#39045) @copyberry
• Restrict MCP HTTP redirects to the configured origin (#39046) @copyberry
• Skip plugin hook loading when hooks are disabled (#39047) @copyberry
• Avoid rendering sub-agent activity twice in the TUI (#39049, #39088) @copyberry
• Prepare the telemetry shutdown worker during initialization (#39050) @copyberry
• Use installed callable apps for TUI mentions (#39051) @copyberry
• Add network policy metadata to environment configuration (#39055) @copyberry
• Verify the pinned zsh manifest in release builds (#39056) @copyberry
• Rate-limit TUI frames from their actual emission time (#39057) @copyberry
• Tag Codex Apps protocol discovery metrics (#39058) @copyberry
• Add desktop app diagnostics to codex doctor (#39060) @copyberry
• Avoid rerendering streamed code fences (#39061) @copyberry
• Render only visible rows in the transcript pager (#39063) @copyberry
• Restrict queued-message editing to its dedicated binding (#39064) @copyberry
• Limit terminal hyperlink layout to the visible viewport (#39065) @copyberry
• Add desktop security enforcement diagnostics (#39067) @copyberry
• Remove skill model delegation support (#39068) @copyberry
• Persist generated images through turn executors (#39072) @copyberry
• Propagate caller metadata to rendezvous connections (#39073) @copyberry
• Add desktop update diagnostics to codex doctor (#39074) @copyberry
• Avoid redundant terminal row clears (#39075) @copyberry
• Build filesystem JSON params only for remote TUI sessions (#39077) @copyberry
• Preserve tracing context for environment resolution (#39078) @copyberry
• Apply user MCP policy to selected executor plugins (#39079) @copyberry
• Bound TUI thread replay buffers by delta size (#39081) @copyberry
• Prompt for project trust in remote TUI workspaces (#39082) @copyberry
• Harden Windows sandbox provisioning against reparse points (#39083) @copyberry
• Preserve filesystem permission path conventions (#39084) @copyberry
• Read plugin authentication state from AuthManager (#39087) @copyberry
• Harden TUI subagent navigation (#39088) @copyberry
• Clarify the external contribution policy (#39089) @copyberry
• Add a command to queue messages for existing sessions (#39092) @copyberry
• Add an agents overview dashboard to the TUI (#39094) @copyberry
• Trace exec-server requests from receipt through completion (#39098) @copyberry
• Avoid redundant terminal size queries during history insertion (#39100) @copyberry
• Update rmcp to 3.1.2 (#39101) @copyberry
• Raise the GPT-5.6 maximum context window (#39102) @copyberry
• Drop capabilities from Linux sandbox processes (#39103) @copyberry
• Make the agents overview an interactive task dashboard (#39112) @copyberry
• Surface interactive requests in realtime conversations (#39113) @copyberry
• Add a dedicated codex agents dashboard command (#39114) @copyberry
• Remove the experimental thread config endpoint (#39115) @copyberry
• Reject lossy legacy permission projections (#39117) @copyberry
• Fail closed on deeply nested command wrappers (#39122) @copyberry
• Validate identifiers in plugin creator workflows (#39131) @copyberry
• Redact auth tokens from app-server response logs (#39141) @copyberry
• Add configurable shortcuts for the agents dashboard (#39142) @copyberry
• Hydrate recommended plugin metadata on selection (#39143) @copyberry
• Persist active permission profiles in turn context (#39145) @copyberry
• Centralize persisted resume settings lookup (#39147) @copyberry
• Update PyPI publish action to v1.14.2 (#39152) @copyberry
• Restore permission profiles when resuming threads (#39153) @copyberry
• Box the TUI future to bound CLI stack usage (#39154) @copyberry
• Prepare Python SDK 0.147.0 stable release (#39155) @copyberry
• Notify clients when Guardian requires strict review (#39157) @copyberry
• Require approval for commands with dynamic shell words (#39159) @copyberry
• Refresh collaboration instructions when their content changes (#39163) @copyberry
• Prevent marketplace identity spoofing (#39165) @copyberry
• Skip empty user messages for automatic idle turns (#39174) @copyberry
• Drop descendant progress updates after remote compaction (#39176) @copyberry
• Scope MCP app resource reads to their originating call (#39187) @copyberry
• Preserve MCP resource origins across compaction (#39192) @copyberry
• Add a symlink-safe reader for sensitive files (#39200) @copyberry
• Reject symbolic links in memory workspaces (#39205) @copyberry
• Add a fail-closed Tree-sitter PowerShell lowerer (#39213) @copyberry
• Prevent custom providers from inheriting ambient auth (#39214) @copyberry
• Reconnect Guardian sampling WebSockets after auth changes (#39220) @copyberry
• Skip redirected external-agent migration destinations ...

Business feature updates

• Collaborate on ChatGPT Sites.

  Site owners can add active members of their Business workspace as editors. Editors can update and save a Site and, after the owner publishes it for the first time, publish later versions to the same Site URL. Owners keep control of sharing, Site settings, analytics, ownership, version restoration, and editor access.

• Use Apple Messages from Codex and ChatGPT Work.

  On Apple silicon Macs, the Apple Messages plugin in the ChatGPT desktop app can read and search iMessage, SMS, and RCS conversations and prepare or send messages through Messages. By default, ChatGPT asks the user to approve the message and recipients before sending. Business workspace admins can disable Apple Messages through the existing Computer Use control.

• Use Computer History in more regions.

  Computer History is now available in the EEA, Switzerland, and the United Kingdom for Business users in the ChatGPT desktop app on macOS. It is off by default, requires Memories, and must be enabled by a workspace admin before a member can opt in.

• Share a read-only snapshot of a Codex chat.

  Workspace share links are limited to authenticated members of the originating workspace, and admins can turn off workspace share links. A snapshot does not update when the original chat changes. Codex redacts known secret patterns, but users should review the snapshot because sensitive paths, diffs, images, or other content may remain.

New features

• Apple Messages: Use the Apple Messages plugin to read and search Messages chats on your Mac and prepare or send messages. It’s available on all plans in the ChatGPT desktop app for macOS. You can use the plugin in ChatGPT Work and Codex. By default, ChatGPT sends messages only after you approve the message and its recipients. See the plugin guide for persistent-approval risks, revocation steps, and the known issue with tasks that disable approval prompts.
• Site co-editing: Where Site collaboration is available, owners can invite active members of the same workspace as editors. Editors can read the Site’s live database data, update the Site, save versions, and publish changes after the owner publishes the Site for the first time. Owners retain control of the audience, settings, analytics, ownership, version restoration, and editor access.
• Editable Site URLs: Where URL editing is available, owners can change an existing Site’s ChatGPT-hosted address without creating another deployment. The previous address redirects to the new URL. Custom domains are a separate, existing feature and aren’t changed by this setting.
• Computer History in Europe: Computer History is now available in the EEA, Switzerland, and the United Kingdom for ChatGPT Pro, Business, and Enterprise users in the ChatGPT desktop app on macOS. It’s off by default and requires Memories. Business and Enterprise administrators must enable access before workspace members can choose to turn it on.
• Shared thread snapshots: On all Codex plans, share a read-only snapshot of a local Codex thread from the ChatGPT desktop app for macOS. The snapshot doesn’t update when the original thread changes. Personal-account links can be opened by anyone with the link; workspace-account links are limited to members of the originating workspace. Codex redacts known secret patterns, but review the shared content because sensitive content may remain. View or revoke links in ChatGPT data controls, under Shared links.
• Unified pinned threads: Keep the same pinned chats across the ChatGPT desktop app and iOS.

Previewing Private Safety Processing, which strengthens safeguards across interactions while remaining compatible with ZDR.

Zero Data Retention gives eligible API customers a clear promise: OpenAI does not retain their prompts or model responses after a request is processed. Customer content is not available to OpenAI personnel for review¹, and enterprise customer data is not used to train our models unless customers explicitly opt-in.

As models take on longer, more complex tasks, some serious risks may only become visible across multiple interactions. Existing ZDR-compatible safety systems evaluate each interaction individually. Today, we’re previewing Private Safety Processing, which is designed to identify patterns across related interactions without giving OpenAI personnel access to the underlying content.

For ZDR deployments, customer content remains on infrastructure the customer controls. We are also developing an option in which content is stored on OpenAI infrastructure, encrypted with keys controlled by the customer. In both cases, automated systems can identify potential misuse and return limited safety signals without exposing the underlying prompts or responses to OpenAI personnel.

Why safety systems need to evolve

The most serious AI safety risks are not always visible in a single interaction. Often, potentially harmful intentions become clear only when multiple interactions are viewed together. Similar risks can arise when bad actors repeatedly probe safeguards, coordinate across accounts, or disguise threats as routine research. Risks can also develop over the course of an agentic task—for example, if a system becomes misaligned with the user’s intent by continuing to act after being told to stop.

As AI systems take on longer and more complex tasks, this broader context becomes increasingly important for distinguishing legitimate activity from misuse and ensuring that AI agents remain within the bounds of their intended authority.

Some recent frontier-model deployments have required customers to allow their AI provider to retain sensitive content for safety monitoring. For many organizations, such requirements conflict with their security obligations or commitments to the people they serve.

Private Safety Processing is designed so we can continue to offer ZDR.

How Private Safety Processing works

Private Safety Processing builds on the automated protections already used in ZDR and other deployments. Existing ZDR-compatible safety systems evaluate interactions individually. Private Safety Processing extends those protections across related interactions, allowing automated systems to identify patterns without OpenAI personnel having access to retained customer content.

Private Safety Processing utilizes customer content regardless of where it is stored—whether in infrastructure customers control (ZDR deployments) or in storage provided by OpenAI. With OpenAI-provided storage, customer content is encrypted using keys controlled by the customer. OpenAI personnel do not have a copy of those keys, so they cannot access the underlying content.

When a risk is identified, OpenAI receives a narrowly defined signal indicating the type of activity involved, similar to our existing safety systems today. That signal can be used to determine whether enforcement is necessary. OpenAI personnel do not receive access to the customer content even when it is flagged.

Customers can investigate alerts and enforcement decisions using information available in their own systems. If they want to appeal, clarify legitimate activity, or support an investigation into verified abuse, they can choose to share relevant information with OpenAI.

Private Safety Processing is currently being tested with early customers. We are sharing this preview now because we’ve heard our customers loud and clear that they need predictability about how their content will be protected as AI systems become more capable.

Privacy and safety built with and for our customers

Our mission is to ensure that artificial general intelligence benefits all of humanity. Collaboration with customers and partners is essential to how we build effective safeguards. As our principles make clear, no AI lab can address emerging risks alone. Private Safety Processing reflects that approach and is being shaped by customers across industries, regions, and company sizes.

The organizations we work with handle some of the most sensitive information in their sectors, including financial records, health data, confidential business plans, and proprietary research. Protecting that information is essential to meeting regulatory obligations, maintaining customer trust, and preserving their competitive advantage.

Their feedback is helping us build stronger safeguards while keeping their information under their control.

“Enterprise AI adoption depends solely on customer control of data, with no direct or derivative use beyond the chosen service. OpenAI’s no-training commitment and ZDR give Glean confidence to build with OpenAI. As models become more capable, OpenAI shows safety can advance without compromising the privacy and control that sustain enterprise trust.”
—Sunil Agrawal, Chief Information Security Officer, Glean

We will continue working with customers on the technical and operational details of our approach. We plan to start rolling out Private Safety Processing, and share a technical white paper, in September. We’ll keep customers informed every step of the way, sharing updates early, explaining what they mean for existing commitments, and providing the time and support customers need to plan ahead.

¹ Like other frontier model providers, OpenAI is required by law⁠ (opens in a new window) to report apparent child sexual abuse material (CSAM). Images flagged for potential CSAM will continue to be retained for manual review and reporting purposes, even in Zero Data Retention deployments, as they are today.

Six months after we began testing ads in the U.S., we’re bringing ChatGPT Ads to 31 European markets.

Next week, ChatGPT Ads will expand to 31 European countries, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, and Austria.

Advertisers will initially be able to access ChatGPT Ads through the OpenAI Ads Solutions team, agency partners, and technology partners. Self-service access through Ads Manager will follow later this summer.

Ads help advance our mission to democratize access to the best AI by supporting free and low-cost access to ChatGPT. As in our existing markets, ads will be shown only to users on the Free and Go plans. Plus, Pro, and Enterprise subscriptions will remain ad-free.

This expansion gives marketers across Europe a new way to reach people while they are actively exploring, comparing, and making decisions.

Where decisions take shape

People come to ChatGPT with goals. They might be planning a trip, choosing software for their business, furnishing a home, learning to understand a new concept or how to start a new hobby. Instead of entering a few keywords, people can explain what they want to accomplish, what matters to them, and the constraints they face.

ChatGPT Ads give marketers an opportunity to reach people across the customer journey as they explore and evaluate their options—and to be useful in the moments that spur discovery, shape decisions, and lead to purchase.

Guided by our ads principles

Protecting the trust people place in ChatGPT is foundational to our approach to advertising.

Guided by OpenAI’s ads principles, we keep conversations private from advertisers and never sell customer data. Ads in ChatGPT are always clearly labeled and separate from ChatGPT’s answers, and advertising does not influence the answers ChatGPT provides. People also have control over ad personalization, with ad-free paid plans available for those who prefer not to see ads.

Our goal is straightforward: ads should be useful to people as they pursue their goals and make decisions—and effective for the businesses that advertise with us.

Building a global ads platform

In February, we began testing ads in ChatGPT with a pilot in the United States. Over the past six months, we have expanded to eight additional markets. The addition of 31 European markets marks our largest expansion to date.

At the same time, we have continued building the ChatGPT Ads platform from the ground up:

• We expanded beyond CPM and CPC bidding to support conversion optimization, helping advertisers optimize toward their business goals.
• We introduced geo-targeting and custom audiences to help advertisers reach more relevant customers.
• We expanded measurement beyond clicks through the OpenAI Pixel, Conversions API, and third-party measurement integrations, giving advertisers a more complete view of how ChatGPT Ads contribute to business outcomes.

Tens of thousands of marketers have now advertised on ChatGPT, and we continue to improve the platform based on what we learn from people and businesses.

We’ve made significant progress in six months, but we are still early in building the full potential of ChatGPT Ads. As we expand into new markets, and develop new formats, optimization tools, and measurement solutions, we’ll remain focused on broadening access to powerful AI, protecting the trust people place in ChatGPT, and helping businesses achieve their goals.

Businesses interested in getting started can sign up at ads.openai.com.

Today’s students will be the first generation to grow up with AI as part of everyday life. For parents and educators, the question isn’t simply whether young people will use AI, but whether they will learn to critically evaluate its outputs, understand its limitations, and use it responsibly.

Right now, there’s a gap between use and understanding. The vast majority of students today are already using AI, while only 16% of high school leaders say⁠ (opens in a new window) all of their students are learning the technical knowledge to understand it in the classroom. Students see this gap too: in a survey⁠ (opens in a new window) by CodeAI, 75% of high school students say understanding AI will be more important for their futures than it is now.

Young people need to understand how AI works, think critically about its outputs, and develop the skills to shape its future with access to tools that prioritize their safety and development.

That’s why OpenAI and CodeAI are announcing a signature partnership to give students and educators access to tools and resources that will help them learn how to use and benefit from AI.

This partnership coincides with the launch of ChatGPT for Teens, a dedicated experience designed with learning at the center to help teens think critically, deepen understanding, and use AI with confidence. It includes built-in protections for teens, including features to promote healthy use and additional controls for parents. It also builds on our ChatGPT for Teachers⁠ (opens in a new window) work to empower teachers and districts to guide AI use and our ongoing work with the American Federation of Teachers to build AI skills. Alongside the ChatGPT for Teens launch, our partnership with CodeAI focuses on the education that surrounds the technology: helping teens understand how AI works, direct it, question it, and create with it—with teachers playing a meaningful role in shaping how AI enters the classroom.

“Parents and educators want young people to be ready for a future shaped by AI—and they are right to expect that to come with care,” said Dr. Allison Mishkin, Head of Child Development, OpenAI.

“Our goal is not simply to help more students use AI, but to help them ask better questions, think critically about the answers they receive, and use these tools responsibly. We look forward to working alongside CodeAI, parents, and educators to learn, improve, and help young people build those skills.”

“Every student should know how AI actually works and be able to question the technology, catch its mistakes, and know when to stop trusting it,” said Karim Meghji, CEO of CodeAI.

“Give them that foundation, and AI becomes less a force shaping their future and instead something they can question, shape, and create on their own.”

How we’ll work together

Over the next year, OpenAI and CodeAI will focus on practical programs that help students understand AI, use it with care, and explore how it can support learning and creativity.

• Creating a continuous feedback loop for responsible AI innovation.

OpenAI and CodeAI will establish a joint advisory council that will bring together experts and leaders on child development, youth public policy, and learning science. The council will provide ongoing guidance on emerging risks, responsible AI practices, and evolving educational needs, helping to shape the future development of ChatGPT for Teens.

• Helping students build foundational AI literacy.

Through the Hour of AI⁠ (opens in a new window), OpenAI and CodeAI will introduce millions of students to the basics of using AI thoughtfully and responsibly.

• Inspiring the next generation of builders.

Through the first-ever Builders Challenge, high school students will have the opportunity to create with AI, receive mentorship from OpenAI team members, and showcase their work on a national stage alongside the teachers who support them.

• Supporting classroom learning.

OpenAI’s experts in AI systems, safety, and responsible use will be a resource that CodeAI can consult as they further develop AI Foundations⁠ (opens in a new window), CodeAI’s free, year-long high school course.

• Making careers with AI more accessible.

Through Career Journeys, students will hear directly from OpenAI researchers, engineers, and leaders about how they work with AI to shape it, helping them see themselves as future builders, creators, and innovators.

Preparing young people for a world shaped by AI is a shared responsibility. Technology companies cannot do it alone. Parents, educators, students, researchers, policymakers, and civil-society organizations all have an important role to play.

Through this partnership, OpenAI and CodeAI will help more young people approach AI with curiosity and confidence—but also with judgment. We want students to learn how to create with these tools without accepting every answer at face value, and to understand that AI should support—not replace—the guidance of teachers, families, and other trusted people in their lives. That is how we can help the next generation become not only capable users of AI, but thoughtful participants in shaping its future.

Built to support learning

We’re introducing ChatGPT for Teens, an experience designed to help teens learn, think critically, deepen understanding, and use AI with confidence. It provides stronger built-in safety protections for teens, including features to promote healthy use and additional controls for parents. If our system estimates someone is under 18 or they state their age is between 13 and 17,⁠ they are automatically placed into ChatGPT for Teens.

This launch builds on the foundational work to support teens on our platform including: Introducing Parental Controls, updating our Under-18 Model Spec, creating our Teen Safety Blueprint, rolling out Age-Prediction and expanding features to support learning and healthy AI use for teens. Across that work, we have been guided by four commitments: put teen safety first, encourage real-world support, treat teens like teens, and be transparent about how our systems should behave. We’re reinforcing and furthering those commitments in ChatGPT for Teens with a focus on learning and even stronger built-in protections.

We believe AI should expand educational opportunity, which is why we designed ChatGPT for Teens to support moments outside the classroom when help is not always available, while keeping teens engaged in an active and collaborative learning process. A teen studying for a test, for example, can ask ChatGPT to explain an unfamiliar concept, work through a practice problem, and check their understanding. If they ask for a quick answer, ChatGPT for Teens can guide them toward the steps and questions that help them work through it themselves. The experience brings together:

• Study Mode, which uses guiding questions and step-by-step support to help teens understand the material.
• New responsible homework reminders, which can recognize when a teen appears to be trying to shortcut an assignment and redirect them toward Study Mode for step-by-step collaborative problem solving.
• Quizzes and Learning Visualizations which give teens more ways to practice, test what they know, and see difficult concepts more clearly to test understanding.
• And with new Study Hours, teens or parents can choose times when Study Mode is on by default, making it easier to turn better study habits into a routine.

“Not every student has someone at home who can help when they get stuck or have questions with their homework. With the right guardrails, AI can help fill that gap. Tools like Responsible Homework Reminder can guide students step by step, helping them understand their mistakes, build confidence, and keep learning—not simply hand them an answer. That kind of individualized support can make a meaningful difference for students who want to succeed but don’t always know how to move forward.”

Racquel Gibson, high school math teacher, Miami, FL

When we introduced Study Mode, it was built with teachers, scientists, and pedagogy experts and drew on longstanding research in learning science. It uses guiding questions, scaffolding, metacognitive prompts, and knowledge checks to help students work through problems step by step instead of simply receiving a solution. That design is consistent with broader research on effective learning techniques⁠ (opens in a new window) showing that active retrieval and self-explanation can support stronger comprehension and longer-term retention than passive review alone. Our own early evaluation of tools like Study Mode has also shown promising gains in student performance.

AI should support learning, not shortcut it, and educators should remain in control of AI experiences in classrooms. In this spirit, ChatGPT for Teens is designed to support learning outside of the classroom while ChatGPT for Teachers continues to provide educators and school districts with secure, institution-managed access to AI and practical support for using it well.

Partnering with CodeAI to prepare teens

To help prepare the first AI generation of students, OpenAI and CodeAI are announcing a signature partnership⁠ to give students and educators access to tools and resources that will help them learn how to use and benefit from AI. The partnership will focus on the education that surrounds the technology: helping teens understand how AI works, direct it, question it, and create with it—with teachers playing a meaningful role in shaping how AI enters the classroom.

“Every student should know how AI actually works and be able to question the technology, catch its mistakes, and know when to stop trusting it. Give them that foundation, and AI becomes less a force shaping their future and instead something they can question, shape, and create on their own.”

Karim Meghji, CEO of CodeAI

What teens are building with AI

Teens are already using AI not only to study, but to build tools, teach themselves new skills, and work on problems in their communities. Those are the kinds of uses ChatGPT for Teens is meant to foster: capable, learning-centered AI that helps young people move from understanding ideas to applying them.

In Virginia, Rushil Kukreja, Arhan Menta, and Nayel Rehman developed WiFind⁠ (opens in a new window), a search-and-rescue system that uses existing Wi-Fi signals to help locate survivors trapped beneath rubble after natural disasters. The project grew out of volunteer work with a local search-and-rescue organization and a simple question about how responders could know where to look first. For the team, AI helped speed up the work of learning from research, prototyping quickly, and turning a promising concept into a tool designed for first responders. Crystal Yang in Texas used ChatGPT to help develop and scale Audemy⁠ (opens in a new window), a platform that creates audio-based educational games for blind students, to 200,000 users and 50 games.

While each teen used ChatGPT in different ways, they all point to the kinds of uses we want to support: learning from complex material, building with guidance, and applying new skills to real problems. Teens should be able to explore that potential with tools that are capable, age-appropriate, and designed with clear safeguards.

Stronger protections, on by default

For teen users, protections are part of the default experience. For users under 18, ChatGPT for Teens applies age-appropriate safeguards designed to reduce exposure to content that may be harmful or developmentally inappropriate, while preserving the ability to learn, create, and explore.

These protections are based on our Under-18 Principles in our Model Spec, which are informed by developmental science and guidance from experts. In practice, that means age-appropriate model protections and product interventions in higher-risk areas such as self-harm, violence, eating disorders, dangerous activities, and explicit sexual or graphic content.

This work also builds on the family tools we introduced with parental controls and notifications. Parents with linked teen accounts can set Quiet Hours, manage selected settings, and receive safety notifications in limited high-risk situations. We are adding additional notifications related to eating disorders, while limiting what is shared and focusing on moments when offline support may matter most.

Supporting healthy, balanced use

We designed ChatGPT for Teens to support real-world relationships and healthy habits, with break reminders that encourage teens to step away, and product cues that consistently identify ChatGPT as AI.

We’re continuing to strengthen ChatGPT’s age-appropriate behavior for users under 18. Our updated under-18 model spec⁠ (opens in a new window) goes beyond blocking romantic or sexualized roleplay: ChatGPT should not use romantic language, encourage emotional dependence, or imply that it has feelings or consciousness. These stronger protections are designed to reinforce healthy, real-world relationships.

We are also adding cues and controls that are meant to reduce risk in context and point teens toward useful tools and protections, including:

• Sensitive-image upload reminders caution teens against sharing private or sensitive images.
• Teen-specific onboarding introduces the learning and safety features available to them from the start.
• And simple customization options, like accent colors and voice variations, help make the experience feel personal without blurring the line between a useful tool and a human relationship.

Ongoing research for safe AI use by teens

These protections are underpinned by ongoing safety research informed by developmental science, expert input, and testing our models’ safeguards in areas where teens may face distinct or heightened risks. This work helps us refine our policies for under-18 users, strengthen product protections, and improve how we measure progress over time.

We are beginning to share more of this work publicly, starting with new under 18 evaluations we’ve added to our system cards⁠ (opens in a new window) that show how our models perform against teen-specific standards in challenging sensitive-content cases. These evaluations cover areas such as self-harm, eating disorders, violence, age-restricted goods and services, and sexual content, and are designed to help advance understanding of model performance on teen safety.

Looking ahead

We will continue to build additional features in order for teens to benefit from AI and measure and publish what we are learning. Teens should be able to use AI to learn, create, and explore. But that access should come with protections that reflect their developmental stage, reinforce real-world relationships, and support healthy use over time. ChatGPT for Teens is our next step in building toward that standard.

We’ve updated the Model Spec, our living document outlining intended model behavior.

This update adds additional clarity around principles on appropriate relational interactions for teens. It also clarifies how assistants should handle false or unsupported premises, removes outdated guidance for pre-reasoning models, adds a new section "Be clear about capabilities and limits", and makes a few other small changes and copy edits.

More information on our principles related to teens can be found in our blog post, and the latest version of the Model Spec is available at model-spec.openai.com.

Work with more of your apps and content

• Apple Messages: Find chats, summarize messages, prepare replies, and send through Messages on your Mac. The plugin is available on all plans in the ChatGPT desktop app for macOS. Use it in ChatGPT Work and Codex, not in regular ChatGPT chats. By default, ChatGPT sends messages only after you approve the message and its recipients.
• Site co-editing: Where available, invite active members of your workspace as editors. Editors can refine the Site and publish updates after its owner publishes it for the first time. Invited editors can read the Site’s live database data; owners retain control of sharing and settings.
• Editable Site URLs: Where available, choose a new ChatGPT-hosted address for an existing Site without redeploying it. The previous address redirects to the new one.
• Computer History in Europe: Use Computer History in the EEA, Switzerland, and the United Kingdom. It remains off by default for ChatGPT Pro, Business, and Enterprise users on macOS. Business and Enterprise administrators must enable access first.
• Shared thread snapshots: Share a read-only snapshot of a local Codex thread from the ChatGPT desktop app for macOS. Personal-account links are viewable by anyone with the link; workspace-account links are limited to the originating workspace. Codex redacts known secret patterns, but review the snapshot before sharing because sensitive content may remain.
• Unified pinned threads: Keep your pinned chats in sync between desktop and iOS.

Find the latest Messages conversation about tomorrow's launch, summarize the open questions, and draft a reply without sending it.

Work with GitLab projects in Codex cloud

GitLab support is available in beta on all ChatGPT plans. Connect a project, create a cloud environment, start tasks from issues or merge requests with @codex, and request one-off or automatic merge request reviews.

The integration runs in Codex cloud, and a managed workspace admin can disable it. GitLab-triggered activity requires permission to configure the applicable webhook. GitLab Self-Managed and GitLab Dedicated connections require workspace admin setup; webhook activity requires GitLab 19.0 or later.

Export public plugin metadata for review

Eligible ChatGPT Enterprise workspace owners and admins can download a CSV of the public plugins visible to their workspace. In Admin > Plugins, select Public, then select the download icon (Export CSV).

The export lists plugin, app, and Chat skill names and descriptions, together with developer, version, date added in UTC, and OpenAI verification metadata. It uses a public-catalog snapshot that can be up to 48 hours old and excludes plugins created for the workspace. The export isn’t available in FedRAMP workspaces.

ChatGPT app experience updates

We’ve made it easier to study with interactive quizzes, manage project memory, and move between typing, dictation, and desktop work.

Learning, Projects, and personalized suggestions

• Practice with interactive quizzes. Ask ChatGPT to quiz you on a topic and answer questions directly in your conversation. Available to all consumer ChatGPT plans and Edu plans on web and mobile.
• Change an existing project’s memory setting. For eligible unshared projects, switch between default and project-only memory without starting a new project. Shared projects remain project-only.
• Get suggestions that fit how you use ChatGPT. Eligible paid users can see homepage suggestions based on their conversation history and connected tools.

Web

• Use Think on Free and Go. Select Think on the web when you want ChatGPT to reason through a harder question.

Android

• Switch between typing and dictation. The dictation button stays available after you add text or attachments, so you can continue building the same message.
• See suggestions that stay in sync. Composer suggestions update as you type or delete text.
• Keep typing with a physical keyboard. Connected keyboards keep the composer focused, preventing interruptions while you write.

Linux desktop

• Use ChatGPT and Codex on Linux. The desktop app is now available globally in public preview on Ubuntu 24.04 LTS and 26.04 LTS, Debian 13, and Fedora 43 and 44.
• Work in the built-in browser or Chrome. Browser actions are available in the Linux app; controlling other desktop apps is not yet supported.

Download ChatGPT for Linux.

Edit memory settings for existing projects

You can now change a project’s memory setting after you create it. Open the project, select the three-dot menu, choose Project settings, and select Default memory or Project-only memory under Memory.

With project-only memory, ChatGPT can use conversations from the same project for context, but it won’t reference memories or conversations from outside the project. Information from the project is also kept out of memory used in chats outside the project. Changes may take a few hours to take effect.

This update is available on all ChatGPT plans. Shared projects continue to use project-only memory and cannot be switched to default memory. ChatGPT Work is not available in projects using project-only memory.

Learn more about Projects in ChatGPT.

Computer History is an optional feature in the ChatGPT macOS app that lets members bring context from selected apps and websites into ChatGPT and Codex. It records interaction events—not screenshots, screen recordings, microphone input, or system audio—and private browsing is not included.

Computer History is off by default for ChatGPT Business. A workspace admin must enable access before members can choose to opt in. Members can pause Computer History, choose included apps and sites, and inspect or delete timeline items. Note that Computer History is not currently available in the EEA, UK or Switzerland.

Computer History is an opt-in feature in the ChatGPT desktop app on macOS that turns activity across apps and websites into memories and a timeline that ChatGPT and Codex can use. Choose which apps and websites contribute, pause collection, and review or delete your history at any time.

Computer History is available to ChatGPT Pro, Business, and Enterprise users. Business and Enterprise administrators must enable access before workspace members can turn it on. Initial availability excludes the European Economic Area (EEA), Switzerland, and the United Kingdom.