Important updates from OpenAI, taken from the company's own pages and rewritten in plain language. Sorted by when we found them, and every entry links to the source.
GPT-5.6 Sol now costs $4 per million input tokens and $20 per million output tokens. That is 20 percent off what you send in and 33 percent off what the model writes out. OpenAI calls it promotional pricing and guarantees it through November 21, 2026.
Why it matters
Sol is the most capable model in the GPT-5.6 family, and the output price is usually the one that dominates the bill. The fact that it is labelled promotional means it can go back up, so the date is worth noting.
Platform
OpenAI API platform
First detected by us
August 22, 2026
Availability
The promotional price runs at least through November 21, 2026.
You can now decide which part of the world an individual request is processed in, by sending it to a domain with a region prefix. The existing requirements still apply: who is eligible, which endpoints and models are supported, and how data retention is controlled.
Why it matters
For European organizations this is the difference between having to set up separate projects per region and being able to steer it case by case. Data that must stay in Europe can be routed there while the rest runs wherever is fastest.
Platform
OpenAI API platform
First detected by us
August 22, 2026
Availability
Requires an API key from a project with Global geography.
Plugin recommendations now rank by which ones people keep using after installing them, not just by install count. ChatGPT also takes better account of your local time when you ask something time-sensitive. Long conversations load faster on the web because messages are fetched in sections, and interactive content starts appearing while it is still being created.
Why it matters
Ranking by continued use is a small but real change: it favours what actually works over what markets itself well. Content appearing as it is built makes the wait visibly productive rather than empty.
Platform
Web and mobile. Desktop is not included in this round.
First detected by us
August 22, 2026
Availability
Which plugins you see depends on plan, region and workspace settings.
A new prompt caching dashboard shows what share of your requests hit the cache, how many reads you get per write, and the split between tokens read from the cache, written to it, or not cached at all. The figures can be filtered by model and service tier.
Why it matters
Prompt caching is one of the cheapest ways to cut costs, but it is invisible: you only notice it on the bill. With the hit rate in front of you, you can see whether the way you build requests actually exploits it.
The gpt-image-2 and gpt-image-2-2026-04-21 models can now produce images with no background. Set background to transparent and choose png or webp as the format. JPEG cannot store transparency and is therefore not supported.
Why it matters
Without this the background had to be removed afterwards with a separate tool, which costs a step and often leaves an edge. Product shots, icons and logos are the obvious use.
Platform
The Images API and the image tool in the Responses API
First detected by us
August 22, 2026
Availability
Preview. Requires png or webp; jpeg does not support it.
A new plugin lets ChatGPT read and search iMessage, SMS and RCS conversations on your Mac, and prepare or send messages through Messages. By default ChatGPT sends only after you approve both the text and the recipients. OpenAI's plugin guide covers the risks of granting persistent approval, how to revoke it, and a known issue with tasks that disable the approval prompt.
Why it matters
This is among the most far-reaching plugins released: it gives a model read access to private message history and the ability to send in your name. The default of per-message approval is what keeps it safe, and it deserves to stay on.
Platform
The desktop app on Apple silicon Macs. Works in ChatGPT Work and Codex, not in regular ChatGPT chats.
The owner of a Site can invite active members of the same workspace as editors. They can read the Site's live database data, update the Site, save versions and publish, once the owner has published it for the first time. The owner keeps control of the audience, settings, analytics and ownership. Owners can also change the ChatGPT-hosted address without redeploying; the previous address redirects to the new one, including sub-paths.
Why it matters
Changing an address without breaking links is the kind of detail that decides whether something can be used for real. Note that invited editors get to read the Site's data, so the invitation is an access decision and not just a division of labour.
Platform
Codex Sites
First detected by us
August 22, 2026
Availability
Where the features are available. URL editing requires Plus or Pro.
You can share a read-only copy of a local Codex thread. The copy does not update when the original changes, and it excludes tool calls and shell input and output. Links from a personal account can be opened by anyone holding the link, while workspace links work only for members of that workspace. Codex redacts known secret patterns, but OpenAI asks you to review the snapshot before sharing.
Why it matters
The caveat is the important part: redaction catches known patterns, not everything. Paths, diffs and images can still reveal more than you intended. Links can be revoked under Shared links in your data controls.
Zero Data Retention is the promise that OpenAI does not keep your prompts or responses once a request has been processed. OpenAI is now previewing Private Safety Processing, a way to spot misuse spanning several interactions without OpenAI staff seeing the content. Automated systems can recognise patterns and emit a narrow signal about the kind of activity involved, while the text itself stays either on infrastructure the customer controls or encrypted with the customer's own key.
Why it matters
Several recent model deployments have required the provider to retain sensitive content for safety monitoring, which collides with confidentiality duties and privacy obligations. This is OpenAI's attempt to avoid that trade, and it is worth following for anyone handling personal or client data.
Platform
OpenAI API platform
First detected by us
August 22, 2026
Availability
Private Safety Processing is a preview. ZDR requires the customer to be eligible.
You can connect a GitLab project to Codex cloud, create an environment for it, start tasks from issues or merge requests by writing @codex, and request merge request reviews either one at a time or automatically. An admin can disable the connector for the whole workspace. Codex cannot complete a review when GitLab omits a collapsed or oversize diff.
Why it matters
GitLab is the standard across many European organizations and in the public sector, where self-hosting is a requirement. Until now they had to choose between their code host and Codex's cloud features.
Platform
Codex cloud
First detected by us
August 22, 2026
Availability
Beta on all ChatGPT plans. Self-managed GitLab needs admin setup, and webhooks need GitLab 19.0.
The /export command writes a whole conversation to Markdown, either to the clipboard or a new file. codex exec fork branches a session, and sessions can be archived or restored from the resume picker. /status and the status line can show the estimated cost of a thread. Amazon Bedrock is now a built-in provider with support for AWS profile, region and GPT-5.6. Hooks can run commands asynchronously and invoke MCP tools.
Why it matters
The cost estimate in the status line is the feature that changes daily use most: it puts the price in front of you while you work rather than afterwards. Bedrock as a built-in provider means companies already on AWS can drop a middle layer.
ChatGPT for Teens is a separate experience with stronger built-in protections and extra controls for parents. It builds on Study Mode, which answers with guiding questions and step-by-step help instead of a finished answer. New are responsible homework reminders, which recognise when a teen appears to be shortcutting an assignment and redirect them to Study Mode, and Study Hours, where the teen or a parent picks times when Study Mode is on by default.
Why it matters
Placement happens automatically from estimated age, so this is not something families have to switch on. For parents, Study Hours is the concrete control: it turns study mode into a habit rather than a choice to be made each time.
Platform
ChatGPT
First detected by us
August 22, 2026
Availability
Users who state an age between 13 and 17, or whom the system estimates to be under 18, are placed here automatically.
ChatGPT Ads is expanding to 31 European countries, with OpenAI naming Norway explicitly alongside Germany, France, Spain, Italy, Sweden, Denmark, the Netherlands and Austria. The rollout was set for the week after the August 18 announcement. Ads are always labelled and visually separated from the answer, they do not influence what ChatGPT replies, and advertisers get no access to your conversations. Advertisers reach it first through OpenAI's own sales team and partners; self-service comes later.
Why it matters
This is the most noticeable change for European ChatGPT users in a while, and it lands on those who do not pay. To avoid ads the options are upgrading to a paid plan, or on the free tier opting out in exchange for fewer daily messages.
Platform
ChatGPT. Norway is among the 31 European countries named.
First detected by us
August 22, 2026
Availability
Ads are shown only to Free and Go users. Plus, Pro and Enterprise stay ad-free.
You can ask ChatGPT to quiz you on a topic and answer the questions directly in the conversation. Projects can switch memory setting after the fact, between default and project-only memory, without starting a new project. Shared projects stay locked to project-only memory. Free and Go users get the Think button on the web for harder questions.
Why it matters
Project-only memory means ChatGPT uses the conversations inside the project as context but keeps them out of everything else. Being able to change the setting afterwards removes an annoying constraint: previously you had to know in advance.
Platform
Web and mobile. The Linux app is announced in the same round.
First detected by us
August 22, 2026
Availability
Quizzes on all consumer plans and Edu. Think on the web for Free and Go.
Ultrafast is a new service tier that runs GPT-5.6 Sol up to 14 times faster than standard processing, at up to 750 output tokens per second. The speed comes from hardware supplied by Cerebras. OpenAI points to uses where seconds count: debugging while an outage is still unfolding, real-time voice support, and research that used to need an overnight run.
Why it matters
Until now, speed has meant picking a smaller model. Ultrafast removes that trade, at least in theory: it is a limited preview, so pricing and broad availability are not settled yet.
Platform
OpenAI API platform
First detected by us
August 22, 2026
Availability
Limited preview for selected customers. Sign up to be notified.
With the Google Drive plugin connected you can browse your files and folders directly from Library, including items shared with you, and pull them into a conversation without uploading them again. Documents, sheets and slides can stay open beside the conversation while you ask ChatGPT to summarize or compare. You can also point at a whole folder and ask it to work across the files. Shared Drives are not included yet.
Why it matters
Uploading was a step that turned files into frozen copies. Now the content stays in Drive, and where supported ChatGPT can update the original directly.
Platform
Web, in both Chat and Work. Mobile follows later.
First detected by us
August 22, 2026
Availability
Plus, Pro, Enterprise, Edu, Healthcare and Business. Requires the Google Drive plugin.
Computer History turns your activity across apps and websites into a searchable timeline and memories that ChatGPT and Codex can use. It records interaction events, not screenshots, screen recordings, microphone input or system audio, and private browsing is excluded. You choose which apps and websites contribute, can pause collection, and can review or delete entries. Business and Enterprise administrators must grant access before members can turn it on.
Why it matters
At launch the EEA, Switzerland and the UK were excluded. As of August 20 they are included, so this now reaches European users. The feature is off by default, and turning it on deserves to be a deliberate choice: a searchable log of everything you do is useful for exactly the same reason it is sensitive.
Platform
The desktop app on macOS
First detected by us
August 22, 2026
Availability
Pro, Business and Enterprise. Off by default, requires Memories. Extended to the EEA, Switzerland and the UK on August 20.
The desktop app is now available for Linux as a .deb or .rpm package, giving access to projects, local files and Codex. The same round lets you import instructions, settings, skills, plugins, projects and recent work from Claude Code, Claude Cowork or Cursor, with automatic syncing if you want it. In Codex CLI, /import does the same from Claude Code or Cursor.
Why it matters
The import is the more interesting half: it lowers the barrier to trying Codex for anyone who has already built up a setup in another tool. Note that the Linux preview lacks Computer Use.
Platform
Ubuntu 24.04 and 26.04, Debian 13, Fedora 43 and 44, on x64 and ARM64.
First detected by us
August 22, 2026
Availability
Public preview, globally. Computer Use is not included yet.
ChatGPT can find available restaurant reservations through OpenTable, Resy and Yelp. You give the place, the time, the party size and any preferences about cuisine, budget or dietary needs, and available times appear in the answer. Pick one and you book it from there.
Why it matters
Outside North America it is OpenTable that applies, since Resy and Yelp are limited to the US and Canada. That decides how useful this actually is: coverage follows OpenTable's own reach.
Platform
Mobile, web and desktop. ChatGPT Work is not included.
First detected by us
August 22, 2026
Availability
All plans. OpenTable globally, Resy in the US, Yelp in the US and Canada.
ChatGPT Business is getting a more expensive seat type with higher usage limits, meant for the people on a team running the heaviest work. OpenAI offered $100 in workspace credits to those who signed up by August 20.
Why it matters
This answers the fact that usage limits bite unevenly: a few people on a team use far more than the rest. Instead of upgrading the whole workspace you can buy higher limits only where they are needed. We discovered this after the sign-up deadline, so the credit is no longer available.
Platform
ChatGPT Business
First detected by us
August 22, 2026
Availability
Signing up by August 20 gave $100 in workspace credits. That deadline has passed.
Daybreak, OpenAI's cybersecurity models, now comes in two access tiers. Daybreak Blue covers ordinary defensive work such as vulnerability discovery, secure code review, incident response and patch validation, using general-purpose models like GPT-5.6 Sol. Daybreak Red requires separate approval and gives access to purpose-trained models such as GPT-5.6 Cyber for authorized penetration testing and red teaming. On August 11 the models also became available through Amazon Bedrock.
Why it matters
The split makes the line explicit: defence is one application, offensive tooling is another. Access is tied to one identity, one workspace, one model and one surface, so it cannot be passed around internally.
Platform
OpenAI API platform and Amazon Bedrock
First detected by us
August 22, 2026
Availability
Requires approval through Trusted Access for Cyber. Daybreak Red needs separate approval.
Plus and Pro users get an updated GPT-5.6 Sol that OpenAI says has more reliable facts and more focused answers. It comes with a slider for how much thought the model puts into a response. Free and Go users move to GPT-5.6 Luna as their default during the week, and from the week after get unlimited text chats plus a Think button for harder questions.
Why it matters
Free users get a markedly better deal: unlimited text and a button for asking for a more considered answer. Limits still apply to file uploads, images and other tools, and OpenAI notes this is subject to abuse guardrails.
Platform
ChatGPT (the chat experience). Work and Codex are unchanged in this release.
First detected by us
August 7, 2026
Availability
Updated GPT-5.6 Sol for Plus and Pro now. GPT-5.6 Luna becomes the default for Free and Go during the week, with unlimited text chats and a Think button from the week after.
Fast mode is the run mode that answers more quickly for a price premium. Until now it could not be used on the very longest requests. Prompts exceeding 272,000 tokens can now run in Fast mode, at up to 2.5 times the speed of standard processing. A token is the small chunk of text a language model counts in.
Why it matters
Anyone sending in large volumes of text, a full body of regulations or a long codebase, no longer has to choose between long context and a short wait. The Fast mode price premium still applies.
The usage and cost dashboards can now be filtered and grouped by API key. The same goes programmatically: both the Usage API and the Costs API accept the key as its own dimension, so the numbers can be pulled out automatically.
Why it matters
A team using one key per project or per environment can finally see who is actually spending the money, instead of reading a single combined total for the whole organization.
OpenAI has built three education plugins: one for K-12 teachers, one for college educators, and one for college students. A plugin here is a ready-made package of apps, role-specific skills and workflows, so the user does not have to construct long instructions on their own. The plugins connect to the material a school already uses, such as documents, course materials and calendars.
Why it matters
A teacher can produce differentiated classroom material from their own files rather than describing everything from scratch each time. This is not something you get on a normal plan: it requires the school or institution to have set up ChatGPT Edu or ChatGPT for Teachers.
Platform
ChatGPT Work and Codex
First detected by us
August 5, 2026
Availability
Through ChatGPT Edu and ChatGPT for Teachers. Not available on ordinary consumer plans.
Paste more than 10,000 characters into the composer and ChatGPT now turns the text into an attachment rather than dropping it straight in. You can pull it back into the message by choosing "Show in text field". The change reached Enterprise and Education on August 4; the other plans already had it.
Why it matters
A long document no longer eats the whole context window, meaning how much text the model can hold in mind at once. The composer also stays readable when you are working with large amounts of text.
Platform
ChatGPT
First detected by us
August 5, 2026
Availability
Rolled out to Enterprise and Education on this date. Every other plan already had it.
The official DALL·E GPT in ChatGPT retires on August 30, 2026. OpenAI asks people to download any images they want to keep before then. Image generation itself is not going away: it continues in ChatGPT Images. GPTs that other users built themselves, with image generation switched on, keep working.
Why it matters
If you have images sitting in conversations with this GPT, there is a deadline to act on. The ability to make images is safe; it is only this one entrance to it that disappears.
Platform
ChatGPT
First detected by us
August 5, 2026
Availability
Applies to the official DALL·E GPT. User-created GPTs with image generation enabled are unaffected.
Codex Security is OpenAI's tool for hunting security holes in code. The plugin can now compare a scan against the previous one, so you see what is new, and you can flag false positives. Findings can be pushed to Linear or GitHub Issues for follow-up. At the same time @openai/codex-security has been made public, so scans can run from the terminal or inside an automated build pipeline.
Why it matters
A team can track whether security holes are actually decreasing over time, rather than reading each scan as a disconnected list. Note that while the package itself is open, running scans still depends on having Codex Security access.
Platform
The Codex Security plugin in the ChatGPT desktop app, plus @openai/codex-security for the terminal and CI
First detected by us
August 5, 2026
Availability
Plugin version 0.1.15 (0.1.14 arrived July 28). The package is public, but running scans requires Codex Security access, and some full scans additionally require Trusted Access for Cyber.
From August 31, 2026, GPT-5.4 and GPT-5.4 mini can no longer be used in Codex by anyone signing in with their ChatGPT account. The models are not going away entirely: they stay on the OpenAI API, and in Codex sessions authenticated with an API key instead. OpenAI recommends gpt-5.6-terra as the replacement for gpt-5.4, and gpt-5.6-luna for gpt-5.4-mini.
Why it matters
Setups still pointing at the old models stop working on that date, and they live in more places than are easy to remember: workspace defaults, saved model settings, managed configurations, custom agents and scheduled tasks. Anyone who needs more time can keep using the models by authenticating Codex with an API key.
Platform
Codex
First detected by us
August 1, 2026
Availability
Applies to Codex users signed in with ChatGPT. The models remain on the OpenAI API and in Codex sessions authenticated with an API key.
The ChatGPT desktop app has a built-in browser, and its address bar can now pull up pages you visited earlier, or search Google when nothing in your history matches. Browsing history is managed in Settings, and ChatGPT can be allowed to search it when a task needs a page you saw before. The Chrome extension lets you mention open tabs or send highlighted page text into a side chat, ask questions about a YouTube video, and start a conversation by right-clicking a web page and choosing "Ask ChatGPT".
Why it matters
Pages you have already read can be brought into a conversation without copying and pasting. Several of these capabilities require the Chrome extension, and letting ChatGPT search your browsing history is something you switch on yourself.
Platform
ChatGPT desktop app and the Chrome extension
First detected by us
July 31, 2026
Availability
Version 26.727. Subscription availability is not specified.
A repository is the folder holding all of a project's code and its history of changes. A multi-folder project now shows every repository it contains and how many lines changed in each. Choosing Review shows the differences across those repositories without switching between separate review views.
Why it matters
A change touching both a website and the system behind it can be read as one piece of work rather than two separate reviews. It requires a project already set up with multiple folders.
Platform
ChatGPT desktop app
First detected by us
July 31, 2026
Availability
Version 26.727. Subscription availability is not specified.
Images ChatGPT has generated can now be opened in an expanded viewer with two ways of looking at them: Focused view and Canvas view. You can add comments across several images, pick the ones you want to work on, and send targeted edits without leaving the conversation. The source does not explain what separates the two views.
Why it matters
Adjusting an image becomes a conversation about that specific image rather than a fresh description from scratch. That makes it easier to fix one detail without losing what already worked.
Platform
ChatGPT desktop app
First detected by us
July 31, 2026
Availability
Version 26.727. Subscription availability is not specified.
The sidebar has a new display called Activity view. It shows which chats you engaged with recently and which ones need attention. It opens by clicking the bell icon or with the Cmd/Ctrl+Opt+U shortcut.
Why it matters
When several tasks are running at once, it is easier to see which one is waiting for an answer from you. OpenAI lists it among the release's smaller improvements rather than as a headline feature.
Platform
ChatGPT desktop app
First detected by us
July 31, 2026
Availability
Version 26.727. Subscription availability is not specified.
Fast mode is a new run mode in the OpenAI API that answers faster for a price premium, replacing the earlier Priority Processing offering. For gpt-5.6-sol, Fast mode delivers up to 2.5 times standard speed at twice the normal price. The change is backward compatible: requests previously tagged priority now automatically use Fast mode, with nothing for developers to change.
Why it matters
Developers who already paid for fast, priority processing get the same benefit under a new name without changing code. The source does not give exact dollar figures, only that the price is double the standard rate.
Platform
OpenAI API platform
First detected by us
July 31, 2026
Availability
Generally available. Requests tagged priority automatically use Fast mode.
OpenAI has cut the price of two models in the GPT-5.6 family: gpt-5.6-luna becomes 80 percent cheaper, and gpt-5.6-terra becomes 20 percent cheaper. OpenAI's own pricing page now lists gpt-5.6-luna at $0.20 per million input tokens and $1.20 per million output tokens, and gpt-5.6-terra at $2 and $12. The cut is not limited to the API: it also carries into how usage in Codex and ChatGPT Work counts against paid subscriptions.
Why it matters
Existing use of these two models becomes cheaper automatically, with nothing to change in code. For anyone working in Codex or ChatGPT Work on a paid plan, the same allowance now stretches further.
OpenAI Presence is a new enterprise product for putting AI agents into production, on the phone and in chat. Each agent gets only the knowledge and system access one specific job requires, such as resolving a billing issue, and the company decides what the agent may do, when it must ask for approval, and when a person takes over. The product does not stop at launch: Codex reads the production sessions and proposes updates the team can test and approve. OpenAI runs its own English-language phone support on it, where it resolves 75 percent of inbound issues without human help.
Why it matters
This is OpenAI selling finished agent operations rather than models alone, a clear step up the value chain. It is also not something you can switch on yourself: Presence ships through a limited program in which OpenAI's own engineers and selected partners carry out the deployment.
Platform
Voice and chat agents
First detected by us
July 31, 2026
Availability
Eligible enterprise customers, through a limited release program. Not self-serve.
ChatGPT for Academic Researchers lets eligible faculty and postdoctoral researchers apply for twelve months in a dedicated ChatGPT workspace for a small, verified team. The workspace supports up to five members, business data protections, and ChatGPT Pro-level usage limits. Applicants use an institutional email address, verify their affiliation, and submit a qualifying research paper. The program does not include OpenAI API credits.
Why it matters
Small research teams can collaborate in a separate workspace rather than mixing the work with other ChatGPT conversations. Admission requires approval, and the program does not replace an API development budget.
Platform
ChatGPT workspace
First detected by us
July 30, 2026
Availability
Eligible faculty and postdoctoral researchers can apply. Geographic availability is not specified.
Sign in with ChatGPT lets a user create or link an account with a supported service using their ChatGPT account. It is available from the ChatGPT plugin directory and participating partner websites. A partner receives only the user's name, email address, and profile picture when available. The user must still approve the access requested by each plugin.
Why it matters
Sign-in and setup can be faster when a service is already connected to ChatGPT or Codex. The beta is limited to selected services, and users should still review each plugin's permissions.
Platform
ChatGPT plugin directory and selected partner websites
First detected by us
July 30, 2026
Availability
Beta rollout to selected plugins and partner sites, starting with Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel
OpenAI released an official Terraform provider for managing resources on the OpenAI API platform. Terraform is a tool for describing and changing technical infrastructure through version-controlled configuration files. The provider can manage projects, users, groups, roles, access, service accounts, certificates, invitations, and project-level rate limits.
Why it matters
Teams can set up and control API access more repeatably, then compare their actual configuration with the intended configuration. It requires an existing Terraform workflow and careful protection of credentials in configuration.
Codex CLI can now name new sessions, pin important threads, and switch between side conversations without closing them. The release also adds Agent Plugins, which are packages of capabilities and configuration, plus more plugin marketplaces. It can fork threads with long histories, connect to remote Code Mode hosts, and discover skills supplied by the execution environment.
Why it matters
It is easier to manage work spread across related tasks and to use more integrations. Some capabilities require compatible model providers, remote hosts, or specific plugin marketplaces.
Platform
Codex CLI
First detected by us
July 29, 2026
Availability
Version 0.146.0. Subscription availability is not specified.
GPT Transcribe is a new model that turns completed audio files and committed speech turns into text. GPT Live Transcribe is intended for live audio and can return text fragments with low latency. Both models can use context, keywords, and language hints to interpret technical terms and multilingual speech more accurately.
Why it matters
Developers can choose between completed transcription and fast text during a conversation, for example in a meeting tool. Results can still be affected by audio quality, language mixing, and the supplied instructions.
Voice conversations now use the voice selected by the user and show usage-limit warnings. The release also improves task continuation, plugin mentions, and skills in the composer, plus tables and themes in built-in visualizations. Several fixes aim to make large change proposals and tasks more stable.
Why it matters
It is easier to continue work on an iPhone after an interruption and use installed plugins while writing. The App Store only lists general fixes, so the specific changes are verified against the Codex changelog.
Platform
iOS
First detected by us
July 29, 2026
Availability
ChatGPT 1.2026.202. Subscription availability is not specified.
ChatGPT can connect to selected digital health records and Apple Health. The information is collected in one health overview, so answers can use the user's own health data as context. OpenAI says the connected data and conversations are not used to train or improve its general AI models, and are not used for ads.
Why it matters
It can make it easier to collect and understand personal health information in one place. The offering is currently limited to signed-in adults in the United States, and ChatGPT is not a replacement for medical care or diagnosis.
Platform
Web and iOS
First detected by us
July 23, 2026
Availability
Free, Go, Plus, and Pro for signed-in users aged 18 or older in the United States
ChatGPT Voice lets users speak with ChatGPT instead of typing. It uses GPT-Live, OpenAI's model for more direct voice conversations, and can start or steer work in other Chat, Work, and Codex tasks. On Mac, Screen context can share an image of the currently open window with ChatGPT.
Why it matters
Users can ask for status, give new instructions, and follow up on long-running work by voice. This reduces the need to open and navigate between tasks manually.
A local project can now contain several related folders. For example, one folder can hold a website, another its backend, and another shared files. The primary folder is the default for new tasks, Git history, and project rules such as AGENTS.md. The other folders can still be read and edited in the same project.
Why it matters
Work spread across several folders can be gathered in one project view. Users do not need to move files or open separate Codex projects for related work.
Platform
ChatGPT desktop app
First detected by us
July 23, 2026
Availability
Version 26.715. Subscription availability is not specified.
Codex CLI is the terminal version of Codex for text-command based work. The release makes it easier to browse long task histories and import configuration from Cursor and Claude Code. It also experiments with Amazon Bedrock, voice input, real-time conversations, and multiple AI agents that can work in parallel on separate parts of a task.
Why it matters
Developers can move an existing AI setup to Codex more easily and manage larger tasks over longer periods. Amazon Bedrock, voice, and multi-agent features are still experimental or require manual activation.
Platform
Codex CLI
First detected by us
July 23, 2026
Availability
Version 0.145.0. Several features are experimental or require manual activation.
Codex tasks on iPhone and iPad can now show Mermaid diagrams, visual flowcharts that explain steps, processes, and relationships. Tasks can also contain forms that users fill in directly. If a user begins writing a message and switches to another task, computer, or workspace, the app attempts to restore the text. The release also improves goals, remote control of Codex tasks, and iPad navigation.
Why it matters
It is easier to understand complex work visually, respond to Codex questions, and follow up on tasks from a phone or iPad. It also reduces the risk of losing an unfinished instruction.
Platform
iOS and iPadOS
First detected by us
July 23, 2026
Availability
ChatGPT 1.2026.195. The App Store records a July 21, 2026 UTC release.
The OpenAI API is the interface developers use to add OpenAI capabilities to apps and services. Organizations and projects can now set a monthly spending cap. When usage reaches a hard spend limit, new API requests are rejected with the technical error code 429. Alerts can be sent before the limit is reached.
Why it matters
A team can prevent API costs from exceeding a defined amount instead of only monitoring usage. Features in an app or service can stop working at the limit, so teams need to plan for that outcome.
Platform
OpenAI API platform
First detected by us
July 23, 2026
Availability
Generally available for organizations and projects
This log starts on Jul 20, 2026. Anything older is deliberately left out, so it is not the vendor's full history.
