held its DevDay on Tuesday 29 September in San Francisco and, according to its official recap, made more than 20 announcements across , , models and the API. We read that recap and the product and documentation pages it links to, and kept what matters to anyone who commissions or builds agents, chatbots and SaaS products. For each item we separate what is available, for whom and at what price, and we give our opinion at the end.
Dots: agents that keep working after you close the chat
The most visible launch is dots, always-on agents inside . They run on , each one has its own cloud computer and browser, and says they can connect to more than 4,000 apps through plugins. You talk to them in , Slack or Teams, and they can start work without being asked: calls this “proactive research” and limits it to read-only tools that cannot send messages or change content. For consequential actions there are custom rules (act, ask for approval or block) plus an automatic review, and tasks such as changing a password always stay with the person.
In Spain there is an important condition. The getting started guide says dots are rolling out to Pro in markets excluding the European Economic Area, Switzerland and the UK, while Business Premium has them in every region where is offered. In Enterprise, Edu and Healthcare they are in beta and stay off until an admin turns them on. A Spanish company that wants to try them will need a business plan.
The idea is very close to Muse, Meta’s agent: an assistant with its own environment that does tasks as well as answering. is betting on WhatsApp’s distribution. places its agent where teams already work (, Slack, Teams) and gives it access to company apps. We think that difference will matter most for businesses, because it decides who configures permissions and who is accountable when the agent gets something wrong.
GPT-6.1 Sol: nearly Astra for a fifth of the price
GPT-6.1 Sol is the update to . In the API it costs 2 dollars per million input , 0.10 per million cached input tokens and 10 per million output tokens. is listed at 10 and 50 dollars in the pricing table, so the one-fifth ratio announced matches the standard rates. It is in the API as gpt-6.1-sol and in Work and for Plus, Pro, Business, Enterprise and Edu. It is not in regular chat yet.
The performance figures come from ’s own evaluations. According to its announcement, on DeepSWE v1.1 (long software engineering tasks) it matches at roughly a fifth of the cost, and on OSWorld 2.0 (computer use), at maximum reasoning effort, it comes within 2.1 points of at roughly a seventh of the cost per task. itself recommends staying on for the hardest scientific research.
For an agent project, the price change matters more than any benchmark. With cached input at 0.10 dollars, an agent that reuses a lot of context between calls becomes much cheaper to run. If you already have a budget built on , it is worth redoing it with Sol and checking which tasks survive the switch. We explain how to do that maths in what it costs to finish a task with an AI agent, and if the names are confusing, in what Astra, Sol, Terra and Luna mean.
Agents API with computer use
The Agents API has been in public beta since 10 September. It offers the harness as a service managed by , which handles context compaction in long sessions, tool search, splitting work across subagents and connecting MCP servers. Code can run in an sandbox, on your own infrastructure or with partners such as Cloudflare, Vercel, E2B or Modal.
DevDay added computer use: the agent drives an OpenAI-hosted browser to test a website, collect information or use an application through its interface. The documentation asks your application to handle each website access request, and sign-in when the task needs it, and to verify the result at the end. According to the recap, the feature is available in the API and also in and Work on Pro 500 and Enterprise. has also worked with Amazon on Bedrock Managed Agents to run these agents inside AWS.
Many of the agents clients ask us for need exactly this: long sessions, their own tools and an environment to run code. That layer used to be built by hand. Now it makes sense to prototype on the Agents API first and keep custom development for cases that need full control over every step or independence from the provider. The harness is open source, which helps avoid being tied to a single provider.
Ultrafast and Pro 500
Ultrafast is the top speed tier. The recap mentions up to 8 times faster generation in (300 tokens per second) and up to 6 times in the API. It is available today for ( says Ultrafast is coming soon), and in the API it costs 60 dollars per million input tokens and 300 per million output tokens, six times the standard rate. The Ultrafast guide adds a detail that matters in Europe: it only supports US data residency and global processing, with no EU regional processing endpoints.
In Work and , Ultrafast is available on Enterprise and on the new Pro 500 plan, at 500 dollars a month, which says includes 25 times the Plus usage allowance. Pro 100 and Pro 200 remain, without Ultrafast.
If your project requires data to be processed in the EU, Ultrafast is ruled out for now. In other cases the extra speed is expensive and only pays off when a person is waiting for the answer in real time.
Codex in the cloud and Codex Security Cloud
Codex can now work from your computer, from your phone or in the cloud. Cloud environments are set up once with the repositories, tools and approved permissions, and each task keeps working while your laptop is asleep. It is available on Plus, Pro, Business, Healthcare, Education and Enterprise. The desktop app adds a code review view where can also take a first automatic pass in the cloud.
Codex Security Cloud scans GitHub repositories on demand or on a schedule, checks new commits and proposes patches. The person reviews the patch before creating the draft pull request. It is available on Pro, Business, Enterprise and Edu.
For maintaining client websites and SaaS products, a recurring security review on every commit looks like the most useful part to us. The patch should still be approved by someone who knows the project.
ChatGPT as a platform: plugins, Space and Pages
Part of DevDay is aimed at anyone with a product. Plugin extensions give a plugin a place in the sidebar, interactive panels and viewers for the file types your product handles. With MCP Events, a plugin can start automations when something happens in the connected app. has also improved plugin ranking and recommendations, in the directory and inside conversations. That organic visibility now sits next to the paid kind we covered in our article on ads and Sponsored Agents in ChatGPT.
Another announcement with an impact on SaaS is Sign in with ChatGPT. Besides signing in with a account, Plus and Pro users can let a participating app use their plan allowance for AI requests, with a weekly limit per app. For some tools this changes who pays for the tokens.
For teams there are Space and Pages: Space replaces Library as the shared home for files, and Pages are documents edited by people, and each person’s dot, which can stay up to date from connected tools. They are available on Pro, Business and Enterprise. Decisions API is in limited preview: it uses the Luna model to answer questions with fixed answers, and is meant for classifying content, routing requests or choosing an agent’s next step.
What people said on X
We went through about fifty recent posts about DevDay. We do not use them as a source of facts, only as a temperature check. @srmarcmayol put it this way for Spain: “Si pagas ChatGPT Pro en España, no tienes dots.” (If you pay for Pro in Spain, you don’t get dots.) ’s guide cited above confirms it. @Daniel0x47 focused on cost: “La prueba será el costo por tarea resuelta.” (The test will be the cost per task solved.) And @technodictAI left advice we agree with: “Before you build on it, check what’s live vs. preview.”
What we would do now
If you already have an agent or chatbot in production, recalculate costs with and test whether its tasks hold up after the model change. If you are building a new one with long sessions or browser use, start with an Agents API prototype before setting up your own infrastructure. And before promising a feature to your team or your clients, check that it is available on your plan and in your region: dots on Pro and Ultrafast with EU data processing are two examples of things that do not reach Europe in the same way.
If you want to work out what fits your case, from a chatbot or agent for your website to an agent that works with your own systems, tell us what you need.