OpenAI DevDay 2026's 25 updates: Dots that do the work themselves, a new model at one-fifth the price, and a $500 Pro

At DevDay yesterday, OpenAI announced 25 major updates in one go: Dots, a personal Agent that keeps following up on your work; the cheaper GPT-6.1 Sol; the faster Ultrafast; and a redesigned Codex cloud experience.

The part most worth trying is how much AI can keep doing after you walk away. Close your laptop and Codex cloud tasks keep running; a Dot can pick up a project without you re-explaining the background every time; and finished documents now have a place to be shared, edited, and handed to a colleague to continue. OpenAI wants the work users hand to AI to carry on beyond the chat window.

The model upgrades are also worth a look. According to OpenAI's evals, GPT-6.1 Sol comes close to Astra on some coding and professional tasks, while its standard input and output prices are only one-fifth of Astra's. Ultrafast takes a different route: pay more, wait less for generation. The first aims to make complex tasks affordable; the second aims to cut waiting time for heavy users.

The batch also includes team documents, meeting notes, plugins, and a new way to bring your ChatGPT subscription allowance into third-party tools. Some of it is already available; some is still in preview or waiting to roll out. Below we go through OpenAI's official list of 25 one by one: what each does, who can use it, what it costs, and which changes are worth trying now.

Proactive intelligence: always-on Agents, cheaper models, faster speed

1. Dots: an AI coworker that never clocks out

Dots are a new always-on Agent in ChatGPT. You give it a goal and set its permissions, and it keeps working on its own, without you prompting it line by line. Everyone can give their own Dot a name.

First, how it works:

• It has its own computer. Each Dot has its own cloud computer and browser, and can write and test code. You can open its computer at any time to see what it's doing, even take over a step and hand it back; unless you explicitly grant access, it can't touch your own laptop.
• It connects to your software. Through plugins, it can connect to 4,000+ apps. You can reach it in ChatGPT (web, mobile, desktop), Slack, and Teams, or call it by voice, and an SMS channel is coming soon. Whichever channel you use, the context carries over.
• It runs on GPT-6 Astra. Astra is OpenAI's most capable model today.

What works best is handing it an ongoing responsibility, for example:

• Developers: watch for requests that keep coming up in customer feedback, scope small improvements and bug fixes on its own, write and test them, and hand you a PR for review with a demo video of the change attached.
• Scientists: when new data arrives, rerun the analysis, chase down unexpected results, update the charts in the paper, and flag what you need to look at.
• Sales leads: check customer requirements against product docs, build proofs of concept for key integrations, and update the proposal as requirements change.
• Content creators: take an interview transcript, pick clips worth cutting, write show notes, and draft social posts for your approval.

On stage, Altman picked a hard example: migrating an app off an old API before it shuts down. Calls to the old API can be scattered across the whole codebase; a Dot follows the dependencies to find everything that needs to change, writes the code, runs the tests, and finally opens a PR for the team to review. Done the old way, this would take several people a good amount of time.

The live demo was given by Holly Li from OpenAI's product team. She used a fictional playlist app to simulate the day before launch. Her Dot is called Dottie:

• By morning, Dottie had already sorted out what she needed to look at: the launch review had been moved up and her calendar updated; last night's tester feedback had been read; design had made a last-minute change to the home page, and the design file was pushed straight to her phone.
• In Slack, a casual @ from a colleague often turns into a whole project. She now forwards the thread straight to Dottie and hands it off with a single "can you take this?" A new habit has even emerged inside OpenAI: when people create a group chat, they add their Dots first.
• She said OpenAI engineers' Dots fix dozens of bugs a day: someone posts a session ID and a user-reported issue in a feedback channel, and a Dot picks it up, investigates, and opens a fix PR.

The demo also stumbled: when she asked Dottie for last night's user test results, it kept replying "still checking"; when she asked it to use Codex on her laptop to turn the new home page into an app in the iPhone simulator, the first attempt errored, and the retry hadn't finished by the time she ran out of time.

Because it acts on its own, control is the most important part of Dots:

• You decide which apps it can access and set rules: what it can do on its own, what it must ask you about first, and what it may never do.
• When nobody is asking it for anything, it does "proactive research" in the background to find ways to help, but this step uses only read-only tools: it can't send messages, change content, or operate your browser or computer.
• Actions that touch your accounts or send information out go through an automatic review first; sensitive things like changing a password always require you in person. It can use your saved passwords to log into websites, but the passwords are never exposed to the model.
• There's an activity view that shows its progress, including what it's doing in the background.

On pricing, your first Dot is included in Pro (including the $100 tier) and Business Premium at no extra charge, with higher limits in the first month after launch. Chatting with it doesn't count against your ChatGPT limits, but tasks it starts for you in Codex or ChatGPT Work are counted as usual. Enterprise has to wait for admins to turn on the beta. At first, Dots can only be created on desktop, and Pro users in the European Economic Area, the UK, and Switzerland are not in the first wave.

Companies get a further form called specialist dots: instead of working for one person, they fill a role in the company, with their own identity, credentials, and system permissions. OpenAI has already tried them internally in procurement, invoice processing, email marketing, customer support, and contract signing. It's now a small enterprise pilot, with OpenAI engineers working directly with companies to define duties, tools, and approval flows; Microsoft will connect it to Agent 365 so companies can manage it with their existing Microsoft tools.

My take: Dots look a lot like Meta's recent Muse, down to the names and cartoon avatars. Many of its capabilities could already be built by people who know what they're doing with Agent frameworks like Codex; what Dots do is package all of that into a product ordinary people can use without setup. You can't tell from one demo whether it's any good. You'd need to use it for a solid week: does it pick up the right things to follow up on, does it remember your corrections, and does it stop to ask you when it should?

2. GPT-6.1 Sol: close to the top model's intelligence at one-fifth the price

OpenAI's current GPT-6 family has three models: Astra is the most capable, Sol sits in the middle, and Luna is the cheapest (our explainer on the GPT-6 Sol and Luna launch). GPT-6.1 Sol is a major upgrade to Sol, focused on agentic coding, computer use, and professional work. Its intelligence is close to Astra's, and its standard input and output prices are only one-fifth of Astra's.

Pricing (API, per million tokens):

Look closely at the price table: Sol's standard price is the same as GPT-6 Sol's from a week ago. What dropped is the cached price: 95% cheaper than standard input, and half the cached price of the previous Sol. The cached price is what you pay when the same context is read again and again. A single Agent task makes dozens of round-trip requests, so that's where most of the bill goes, and people building Agents should look at the cached price first.

In the evals OpenAI published, Sol compares with rivals like this:

• DeepSWE (software engineering tasks in real codebases): matches Astra at about one-fifth the cost, 6.4 percentage points above the previous Sol's best score.
• GDP.pdf (reading complex PDFs to answer professional questions): beats Opus 5.5 at every reasoning level, at less than half its cost per task.
• AutomationBench (running multi-step business workflows with 47 tools): at medium reasoning effort, 2.2 percentage points above Opus 5.5 at about one-third the cost.
• OSWorld 2.0 (long-horizon computer use): within 2.1 points of Astra at about one-seventh the cost per task.
• Terminal-Bench Science (research tasks): an average of $5.47 per task, versus $23.21 for Opus 5.5 and $23.80 for Astra; but Astra still has the top score (68.1%), so the hardest research work should still go to Astra.
• Factual errors: on hard questions picked because they're easy to get wrong, the share of answers containing factual errors at low reasoning effort fell from 11.4% to 7.7%; at the highest setting, Sol is at 4.1% and Astra at 4.0%, nearly even.
• Honesty: when the search tool breaks, the model should tell the user rather than make up an answer. The share of cases where it didn't say so is 2.1% for Sol, 4.9% for the previous Sol, 1.5% for Astra, and as high as 28.7% for Luna.

The raw DeepSWE data also shows something more useful for users: turning reasoning effort higher isn't always better.

At "Medium," the new Sol spends $0.42 and scores slightly above Astra at the same setting, which costs $3.08. Its best score comes at "High"; push it to Max and the cost more than doubles while the score drops by more than 3 points. Astra is the same: Max costs more than "Extra high" and scores lower. So when choosing reasoning effort, try "Medium" and "High" on your own tasks first rather than defaulting to the maximum.

On an external independent composite leaderboard, Sol scores 52 and Astra 53, and Sol's cost per task is about 22% of Astra's, so the claim of "close to Astra at about one-fifth the price" holds up. First place still goes to Anthropic's Opus 5.5 (58 points). Sol competes on value.

Specs: 1.05M-token context window, 128K-token maximum output, text and image input, five reasoning levels. Prices double when a single input exceeds 272K tokens; Batch and Flex are half price.

Sol isn't in the chat window yet. It's available in ChatGPT Work and Codex (Plus, Pro, Business, Enterprise, Edu), and developers can use the API model name gpt-6.1-sol.

3. Ultrafast: pay more for faster output

Ultrafast is a paid speed tier. In Codex, token generation is up to 8x faster, about 300 tokens per second; in the API, it's 6 to 7x faster.

An Agent takes dozens of small steps to get one thing done, and it waits on the model at every step, so a faster model cuts your wait by the same margin. On stage, Altman showed a side-by-side comparison with the same prompt, "Build a white rocket in DevDay colors with a big porthole, and launch it": on the Ultrafast side the rocket had already lifted off, while the standard side was nowhere near finished building it.

The cost is money. In the API, Ultrafast is 6x the standard price: Astra Ultrafast is $60 per million input tokens and $300 per million output tokens. In ChatGPT plans, using Ultrafast burns through your limits 8x as fast. It runs on Cerebras chips.

What's available now is Astra Ultrafast: it's live in the API, and in ChatGPT Work and Codex only Pro 500 and Enterprise can use it. Sol Ultrafast arrives in the next few days. Its price will be roughly what standard Astra used to cost, yet it gets you near-Astra intelligence at up to 8x the speed, which may be the best-value combination for developers in the whole launch.

It's worth paying for when a person is sitting there waiting on the result, such as pair-programming with AI in Codex. For batch jobs nobody is waiting on, standard-speed Sol plus caching is much cheaper.

4. Private Intelligence: for companies wary of handing over data

For customers like hospitals, banks, and law firms, data is the biggest worry. Private Intelligence is OpenAI's answer for them, in two parts:

• Zero Data Retention (ZDR) with Private Safety Processing. It resolves a conflict: a company that turns on zero data retention doesn't want OpenAI to keep prompts and responses, yet OpenAI still has to do safety reviews to prevent model abuse. The approach is to move the records to the customer's side: conversations flagged by the safety system are encrypted and written to the customer's own cloud storage, and OpenAI keeps only an index, not the content. When a review is needed, only a hardware environment that bars human access can decrypt the records; it reviews them automatically and outputs only a safety signal in a fixed format.

  What "zero data retention" means precisely: OpenAI doesn't keep your content; the safety records sit in your own storage for 30 days, managed by you. Add a layer of the company's own keys, and once those are revoked, nobody can decrypt the records.

• Private inference (preview). Uses confidential computing so that data stays under verifiable protection even at the moment the model generates a response. Preview starts this fall.

Altman said the system was designed together with large customers such as Cisco, Databricks, and Snowflake.

Building with Codex and the API: moving Agents to the cloud and giving developers more parts

Codex is OpenAI's coding Agent. This section has 7 items: 4 add capabilities to Codex itself, and 3 are API building blocks for developers.

5. Codex Cloud: close your laptop and it keeps working

As soon as Altman took the stage, he checked off the three Codex requests users had asked for most loudly on X: a Codex app for Linux, projects that span multiple folders, and Codex on mobile.

The bigger change is where it runs. Codex now has three modes: run on your computer, run on your computer controlled from your phone, or run directly in the cloud. In the cloud, tasks no longer depend on your laptop staying open.

What's new is reusable cloud development environments: set one up once, and every later task starts right away. Teams can also share one approved set of settings and permissions.

The film shows several ways to use the new environments:

• Keeps running with the lid closed: tasks run in the cloud and don't stop when you close your laptop. You can add to a running task from your phone, or even start a session by voice with your earbuds in, and it notifies you when it's done.
• Conversational setup: when creating an environment, hand it the repo and it works out what the repo needs on its own: network access, environment variables, secrets.
• Set up once, reuse often: once configured, save it as a named environment snapshot and launch multiple independent tasks from it at the same time. When shared with a team, members can swap in their own secrets.
• Wrapping up: when done, it can open a PR directly or go through the new code review flow (item 7).

On stage, Romain Huet demonstrated a handoff between local and cloud: he created a cloud task with just one line, "Rewrite the entire backend in Rust," went on to demo other things, and came back later to check the result.

Available on Plus, Pro, Business, Enterprise, Edu, and the healthcare plan.

6. A refreshed Codex CLI: talk to it, and manage multiple tasks at once

The CLI is Codex on the command line. What's changed:

• Voice conversation: start tasks and correct course just by speaking.
• /agents view: split work across multiple Agents and see the progress of each in one interface.
• Everyday polish: easier prompt editing, resuming previous sessions, built-in Git worktrees (so multiple tasks run in separate copies of the code without interfering), and long sessions that are easier to read.

The /agents view deals with the mess of multitasking: when AI is doing several things at once, it's easy to lose track of which one is waiting for a reply, which is editing files, and which is done and waiting for review. The more tasks you run, the more useful this view gets.

On stage, Romain started from an empty project and had the CLI build an app that randomly picks three audience members to win tickets to the next DevDay. The page appeared within seconds; he then said "make it six," the page updated instantly, and six audience members were drawn on the spot. He had planned to demo this by voice, but voice wouldn't start that day, so he switched to typing. Available on all plans.

7. Code review: Codex takes a first pass while you're away

The ChatGPT desktop app now has a dedicated code review interface: read a summary of the changes first, then drill into specific diffs, and ask Codex directly when something's unclear. Turn on automatic review, and Codex reviews in the cloud first when you're away from your computer. It supports GitHub pull requests and GitLab merge requests, and is available on all plans.

8. Codex Security Cloud: AI that watches your code for vulnerabilities

Security Cloud is aimed at security teams. It has Codex continuously watch code repositories in the cloud for vulnerabilities.

• It can watch every new commit, or scan an entire GitHub repo on demand or on a schedule.
• When it finds a problem, it investigates, deduplicates, and prepares a fix patch on its own, all in the cloud, so it keeps running with your laptop closed.
• It comes with the defense-oriented Daybreak Blue model from OpenAI's cybersecurity program, with no separate application needed.

Scanning itself isn't hard. The time goes into what comes after: judging whether it's a real problem, deduplicating, locating it, writing a patch. Security Cloud does that round for security teams first. It's currently a research preview for Pro, Business, Enterprise, and Edu, offered as a plugin in the Codex desktop app and on the web.

9. Decisions API: a fast model that only answers multiple-choice questions

The problem: many apps involve a kind of "small decision": which team should this ticket go to? What category does this content fall into? Which tool should the Agent call next? Today you usually call a large model, force it to output only a set format, then write code to check the format and retry on errors. It's slow and roundabout.

The approach: you define a question and a few fixed answers, give it context in text or images, and it returns one of the options directly. Under the hood is GPT-6 Luna, the cheapest model.

Routing the same 10,000 customer service requests into billing, technical, and sales, the regular Responses API takes about 1.6 seconds per request, while the Decisions API takes about 150 milliseconds, roughly 10x faster.

On stage, they also demoed using it to operate a computer and book a flight. The footage wasn't sped up, and each step flashed by almost instantly. Another use is robotics: Romain ended by bringing out MicroDuck, an unreleased programmable little duck robot from Hugging Face, using Astra for vision, GPT Image 2.5 for drawing, and GPT Live 1 for real-time voice. He said what his robotics friends are most excited about is using the image-capable Decisions API to let robots act on what they see in near real time.

This kind of model is sometimes called System One (fast thinking), and it follows a similar idea to Jev, which TypeSafe launched recently: let a strong model think about the big picture, and let a fast decision model pick each step. When you use it, remember to include a "needs human review" option so ambiguous inputs don't get forced into the wrong category. It's in limited preview now, with general availability within days.

10. Agents API adds computer use: let your Agent operate software directly

The Agents API is OpenAI's cloud Agent hosting service. It launched earlier; what's new this time is computer use.

Building an Agent that can work for long stretches takes more than a model. You also have to manage context, call tools, coordinate multiple sub-Agents, and have a runtime that can run for days. This whole setup is called a "harness," and the Agents API hosts the one Codex uses for you:

• One API call creates a cloud Agent: just specify the task, model, tools, and runtime environment.
• Pick your environment: an OpenAI-hosted sandbox, your own infrastructure, or environments from partners such as Cloudflare, Vercel, and Modal.
• Built-in capabilities: automatic compaction when context is nearly full, tools loaded on demand, and multiple sub-Agents working in parallel.
• Billing: the Agents API itself is free; you pay only for tokens and tool usage.

The new computer use lets your Agent click, type, and operate software interfaces the way a person does. Many real-world systems have no usable interface at all; you can only open a web page, check the status, and click buttons. Computer use fills that gap. It's available through the API; Pro 500 and Enterprise users can also use it in Codex and ChatGPT Work.

Altman also announced that Codex Harness is now open source. It powers Codex and ChatGPT Work, and now Dots too, and developers can take it directly to build their own Agents.

Tejal Patwardhan, who leads post-training research at OpenAI, shared two sets of numbers on computer use. On speed, they had a model repeatedly search in a loop for latency-cutting improvements and merged them into the production harness; computer-use latency improved by more than 2x, and that's already live. On safety, in a computer-use safety stress test, the rate of behavior deviating from user intent was 9.5% for Claude Fable 5.1, 6.1% for Opus 5.5, and 2.4% for GPT-6 Astra.

She also gave a number on research progress: for research tasks that take a person about a day, models could barely complete any at the start of the year; by mid-year, more than a third could be completed without human intervention.

Romain showed off these capabilities in several demos: Astra took over a 3D space game it had never seen through the browser, clicking start, picking a ship, and flying it on its own. He also took an AppShot of an app in a phone simulator and said "review my app across all screen sizes," and Codex went into the simulator on its own, opened the app, tried each feature, and took screenshots.

The API as a whole got faster too: tool calls are 30%+ faster, time to first token is 45%+ faster, and Responses API availability is 99.9%+.

11. Amazon Bedrock Managed Agents: running entirely inside AWS

For companies whose data and procurement sit in AWS, this item solves the "we want OpenAI's Agents but don't want to leave AWS" problem. Bedrock Managed Agents build on the core capabilities of the Agents API, add AWS customization and integrations, and run on the customer's own AWS resources, right next to existing apps and data. AWS customers can also use GPT-6 Astra, Ultrafast, Codex, and ChatGPT Work.

Customizing ChatGPT with plugins: turning ChatGPT into a front door for your product

From GPTs in 2023 to apps in ChatGPT later on, it's all now called plugins. ChatGPT, with 1.2 billion weekly users, has opened up to developers the same platform it uses to build its own features. All 4 items in this section are for developers.

12. Plugin extensions: build a full app inside ChatGPT

Plugins used to be more like "a tool ChatGPT can call"; now they can grow into "a full app inside ChatGPT." Developers can:

• Give a plugin its own entry point in the sidebar;
• Build an interactive panel that users can work in while chatting;
• Build viewers for the file types their product supports.

Altman gave two examples. One is OpenAI's own new meetings app, built with plugin extensions: you see upcoming meetings in chat, click "Take notes in ChatGPT," and the notes page becomes a ChatGPT Space where colleagues and Dots follow up on action items. The other is Figma: open a design file in ChatGPT, read the team's comments, have ChatGPT make a revision, with the design and the conversation side by side. Available on all plans.

13. Easier creation, submission, and discovery

• Creation: a plugin builder helps you put it together.
• Submission: the submission flow has been rebuilt. You can see review progress, see what needs changing, and request a human re-review, and plugin updates no longer require resubmitting from scratch.
• Discovery: besides search in the plugin directory, when users ask a question, ChatGPT judges whether your plugin can help and can connect it on the spot.
• Users decide: users choose which plugins to use, and approve each plugin's permissions one by one.

For plugin developers, in-conversation recommendations matter most: questions from 1.2 billion people a week could all become distribution channels. It also looks a lot like a sponsored slot in a search engine, and how it's ranked and whether it will cost money remain to be seen. Available on all plans.

14. Sites with plugins: teams log into the same app with their own permissions

Sites is ChatGPT's website-building feature. In the few months since launch, users have built millions of sites. These sites can now connect to ChatGPT plugins: team members log into the same app with their own accounts, use the data and permissions they've connected themselves, and see different content accordingly. Finance dashboards, event booking, project tracking, budget planning, and multiplayer mini-games can all be built this way. Available on Business, Enterprise, the healthcare plan, and Education.

15. Event-driven automation with MCP Events: it springs into action when something happens

MCP is an open protocol for connecting AI to external tools. ChatGPT now supports MCP's event mechanism: when something happens in a connected app, it can automatically trigger a plugin workflow.

It works in the opposite direction from regular tool calls. With regular calls, the AI asks the external system "anything new?" and has to keep going back to ask; with events, the external system pushes a message as soon as something changes. The flow is roughly: your service declares which events it supports (say, "someone posted a new comment"); the user tells ChatGPT what to watch for and what to do when it happens; ChatGPT subscribes through your MCP server; then whenever a matching change occurs, your service pushes the event over, and ChatGPT carries on in the original conversation according to the instructions.

For example, have ChatGPT watch a project board, and whenever a new task appears, whether or not you're around, it reads the linked docs and drafts a plan. Available on all plans.

Humans and AI working together: a shared workspace

The 7 items in this section can be read together. What they build is "an office where people, ChatGPT, and Dots work side by side."

16. ChatGPT Space: a place shared by the team and AI

Bring your team into a dedicated space, and every member, ChatGPT included, can build on shared goals and past work without re-explaining the background each time. You can ask ChatGPT to turn a conversation into a page or slides; it first clarifies what you need, then drafts it in front of you, and you can edit it yourself, have it revise, or pull in colleagues to edit together, all without switching tools. It replaces the old Library; think of it as "a shared drive plus a group chat, except some members are AI."

Available on Pro, Business, and Enterprise. You can create and edit on web and desktop; on mobile it's view-only for now. Plus isn't on the list.

17. Pages: a new kind of document for people and Agents to edit together

Pages are suited to project plans, reports, and visual proposals. Inside one, you can work with ChatGPT or your Dot to generate charts, interactive dashboards, and images; colleagues can comment or edit directly, then @ your Dot or ChatGPT to revise content or take the next step.

In her demo, Holly @-mentioned two colleagues and Dottie and said "chart, please," and a user feedback chart filterable by feedback type appeared on the page, updated hourly by Dottie. Pages can also be set to update automatically from connected tools, such as a to-do list that changes as work progresses.

Before your team starts using it, pay attention to permissions: sharing a page doesn't share your private chats and memory, and doesn't hand out access to external files. But once ChatGPT excerpts or summarizes that content into a shared page, everyone who can see the page can see it. It's best to set ground rules first for what can go into shared pages. Available on Pro, Business, and Enterprise.

18. Collaborative slides: several people plus AI editing one deck at once

A new presentation editor:

• Natively editable charts, plus text, shapes, and images;
• Turn a conversation straight into slides, or start from your own template and give ChatGPT feedback as you go;
• Multiple people can edit and comment at the same time, or mark spots they want ChatGPT to change;
• When finished, present directly in ChatGPT, or export to PowerPoint or Google Slides to keep editing.

Altman described it as "a kind of slide deck that's easy for Agents to read and edit." What it cares more about is keeping content, layout, and feedback in one place as several people go back and forth. Rolling out in the coming weeks for Pro, Business, and Enterprise.

19. Teams and shared tasks: hand repetitive work to ChatGPT

Pull colleagues into a team in ChatGPT and you can share pages, presentations, and spreadsheets. The more useful part is team tasks: take a weekly project status report, for example. You describe what you need, and ChatGPT gathers information and takes actions with the connected tools. It can run on a schedule or be triggered by changes, such as a new email or a new Slack message. Team members can edit the instructions together so the task shifts with the team's priorities. It runs under the team's service account rather than borrowing all of one person's permissions. Available on Business and Enterprise.

20. @ChatGPT in Slack and Teams

@ChatGPT in a channel, thread, or DM and it can turn a discussion into a project brief, investigate a bug and prepare a fix, or send a customer update every morning. It can use tools connected by admins, and your own tools within the scope you authorize. One detail: colleagues without a ChatGPT account can also add context and help refine the result in the same conversation. Available on Business and Enterprise.

21. Meetings plugin: notes and action items ready when the meeting ends

Record meeting audio with the Meetings plugin, and the notes are saved to ChatGPT Space, either private to you or shared with the team. ChatGPT draws on the context of your past work together to pull out the points, decisions, and next steps most relevant to you, and can help with action items, such as updating the project plan after a standup based on new decisions and blockers.

The recording is deleted once the notes are generated and can't be played back, so don't count on it where you need to keep the original audio. It's currently a macOS desktop beta for Pro and Business, with Enterprise to follow.

22. Shareable profiles: a business card for your plugins and sites

A profile page showcases the sites and plugins you've built so others can find and reuse them; team members can also discover shared "skills," which are shared only within the workspace. For people who build small tools, this is much easier to find than links scattered across group chats. Coming soon to Enterprise, Edu, and the healthcare plan; other plans can already use it.

Doing more with your ChatGPT subscription: accounts, plans, procurement

23. Sign in with ChatGPT: one account that goes into other products

Log into partner products with your ChatGPT account, no new password needed. The partner only gets your name, email, and profile picture; your chat history and memory aren't shared.

Beyond login, there's usage: Plus and Pro users can spend their ChatGPT plan usage directly in partner apps, and you decide the maximum each app can use. There are 16 launch partners, including Devin, Notion, Vercel, Lovable, Warp, Amp, Conductor, Kilo, T3, OpenClaw, and Dactyl. Note that this comes out of your existing allowance; connecting another app doesn't give you any extra.

Altman admitted on stage that this kind of thing has been tried before, "with mixed results." The benefit for developers this time: inference costs are paid by the user's subscription, so you no longer have to front token costs for heavy users. The cost: your login, your users' allowance, and your cost structure all rest on OpenAI's account system.

24. A new Pro tier: $500 a month

OpenAI launched Pro 500 at $500 a month: 25x the usage of Plus, no 5-hour limit, an allowance that can also be spent in partner apps via "Sign in with ChatGPT," and it's the only tier that can use Ultrafast in ChatGPT and Codex; buying extra usage on the $100 or $200 tiers won't unlock it. Pro now has three tiers: $100, $200, and $500 a month.

The $200 tier changed too: it's open for new subscriptions again, but new subscriptions get less usage. In ChatGPT Work and Codex it drops from 20x Plus to 10x, and GPT-6 Pro messages in chat drop from 200 to 100 a week; existing subscribers keep their old allowance during a transition period. Heavy users on the $200 tier should redo the math on their usage.

Whether to upgrade depends on whether limits and waiting hold you up every day. For people working in Codex all day, $500 for 25x usage plus Ultrafast may be worth it; for occasional users, the extra money won't buy a matching improvement.

25. OpenAI Marketplace: buy other vendors' software with your OpenAI contract commitment

Eligible enterprise customers can use part of the spend they've already committed to OpenAI to buy partner software, without going through procurement separately for each vendor. There are 32 launch partners: Adobe and Figma in creative; Sierra, Decagon, HubSpot, Salesforce, and ServiceNow in customer experience; Harvey and Legora in legal; Palo Alto Networks and CrowdStrike in security; plus Baseten, which provides open-source models. It's currently in beta, and companies can register interest now.

One table: who can use what right now

Putting the 25 together

Sorted by "which kind of work cost it lowers," the 25 items fall into five groups:

• The cost of ongoing follow-up: Dots. Before, you had to assign the task and also keep nudging the AI to carry on. What Dots aim to remove is the attention that "keeping an eye on it" takes up.
• The price of capability and the price of waiting: GPT-6.1 Sol and Ultrafast. One lets more complex tasks go to a cheaper model; the other lets urgent tasks pay for speed. These are two separate bills and should be calculated separately.
• How much tasks depend on a person being there: Codex Cloud, MCP Events, team tasks. With a place to run, events to trigger them, and an identity the team manages together, automation can go from one-off scripts to processes that run long-term.
• How finished work gets handed off: ChatGPT Space, Pages, plugin extensions. AI output moves out of chat logs and becomes documents that can be edited, shared, and acted on further. This group is closest to the everyday office worker.
• Paths for distribution and procurement: Sign in with ChatGPT, OpenAI Marketplace. The first lets personal subscription allowances travel outside ChatGPT; the second lets corporate procurement budgets flow to partner software. Neither has to do with model capability; they affect where people will use and buy AI tools going forward.

The collaboration items look a bit scattered on their own, but put them into one handoff flow and they fit together:

A complete flow looks like this: a meeting produces action items, the action items go into a page, the page goes into a shared space, team tasks or a Dot watch for later changes, and the results come back to where everyone is discussing.

A few open questions

• The live demos kept stalling. A Dot stuck on "still checking," voice failing to start twice: a clear contrast with the smooth day in the promo film. Whether products like Dots are any good depends on how many things you stopped chasing, and how much rework you picked up, after a month of use.
• Whether $500 is expensive depends on whether you can show a return. $6,000 a year per seat isn't outrageous for a developer tool, provided you can say how much time it saves. The value of the $500 tier is heavily concentrated in Ultrafast, and Ultrafast only shows its worth when someone is sitting and waiting.
• Will in-conversation plugin recommendations turn into ad slots? The moment when 1.2 billion people ask questions is the best distribution channel there is. How recommendations are ranked and who comes first depends on what happens next.
• Dots have run into Muse. Both companies are building personal Agents that "have a name, have an avatar, and are always on." OpenAI's advantage is that Codex, plugins, and ChatGPT's user base sit behind it; in the end, the contest is over who first gets ordinary people comfortable handing over a responsibility.