I watched all 20+ OpenAI DevDay 2026 launches. Here are the 5 that change what I build

OpenAI DevDay 2026 ran on 29 September at Fort Mason in San Francisco. Sam Altman opened at 10:00 AM Pacific, with Romain Huet, Tejal Patwardhan and Holly Li following him on stage. The opening slide put the company at 1.2 billion weekly ChatGPT users, more than 40 model launches and 2.5 million businesses. More than twenty things shipped in about an hour.

I watched the keynote and went through the official recap afterwards. This is what actually changes the work if you build agents, rather than what made the headlines.

dots is the headline, and it is a real product

The leaks said the always-on assistant would be called "o". The Verge reported a codename of Aeon. Both were wrong. It is called dots, and it is OpenAI's answer to Meta Muse and Grok Bot.

A dot is an always-on agent that lives inside ChatGPT. Each one runs on GPT-6 Astra, gets its own cloud computer and its own browser, and works toward a goal you hand it around the clock instead of turn by turn. The detail that matters most to me is continuity: a dot keeps context across surfaces, so a project you open in ChatGPT carries on inside a Slack thread with no recap and no re-prompting. It reaches more than 4,000 applications through plugins, and with your permission it can drive your own laptop in addition to its cloud machine.

Availability is Pro and Business Premium, with beta access for Enterprise, Education and Healthcare where an administrator turns it on. It is off by default.

GPT-6.1 Sol is the quieter win

GPT-6 Sol shipped on 22 September. One week later it already has a successor. GPT-6.1 Sol is aimed at agentic coding, computer use and professional work, and OpenAI says it delivers near-Astra intelligence at one fifth of Astra's standard input and output token pricing.

If you are running agent loops where token cost decides whether the architecture survives contact with a real workload, this is the more consequential launch of the day. Astra is the model that gets talked about. Sol is the model most of us will actually be able to afford to leave running.

The ultrafast service tier that appeared in OpenAI's public OpenAPI spec on 25 September also turned out to be real. Astra Ultrafast is available in ChatGPT, Codex and the API.

Codex moved fully into the cloud

Codex now runs entirely in the cloud, which means your machine no longer has to stay awake for a long task to finish. Reusable development environments let a task start immediately with approved settings and permissions, and those environments are shared across a team rather than rebuilt per developer. It is on Plus, Pro, Business, Healthcare, Education and Enterprise.

Around that came a refreshed Codex CLI, a code review surface that discusses individual changes, and Codex Security Cloud with a findings dashboard. Codex mobile appeared as a ticked item on a slide with no date attached to it.

The API news, including one deadline you should write down

There is a new Decisions API, and the Agents API now has computer use, multi-agent capability, tool search, tool calling and context compaction, with OpenAI running the infrastructure underneath. Amazon is launching Bedrock Managed Agents powered by OpenAI so those agents run natively inside AWS.

The part worth acting on: Agent Builder, the workflow builder launched at DevDay 2025, loses access on 30 November 2026. If you wired anything through it, the migration path is the Agents SDK and Workspace Agents, and you have two months.

Plugin extensions turn ChatGPT into a distribution channel

This is the platform move. Plugins can now be, in Altman's words on stage, essentially entire applications that feel native to ChatGPT. They live in the sidebar as interactive panels beside the conversation, they carry file-type viewers, and OpenAI distributes them. It is available on all plans, and Canva, Figma and Shopify were the demo names.

Alongside that: improved plugin creation, submission and discovery, Sites can host plugins, and MCP events wire plugins into automations. ChatGPT Space adds a shared workspace where a team and its dots work on shared knowledge, with collaborative Pages, collaborative presentations, shared team tasks, @chatgpt in Slack and Microsoft Teams, and a Meetings plugin that takes actions off a recap.

Sign in with ChatGPT is the sleeper feature. Your plan allowance now travels to other tools across 18 partners, including Cognition's Devin, Notion, Vercel, T3, OpenClaw and Dactyl, so eligible usage counts against your ChatGPT limits instead of arriving as a separate bill.

There is also a new Pro tier. The keynote slide reads Pro 500, and a 500 dollar per month price has been widely reported, though I have not seen it on an OpenAI pricing page yet, so treat the number as reported rather than confirmed.

What did not ship, and the context around all of it

No GPT-6 Cyber, no hardware device, and as noted, no product called o or Aeon. Three of the loudest pre-event rumours went unmentioned entirely.

GPT-6.1 Astra stayed shelved. OpenAI pulled it the day before the conference after deciding it did not meet the company's safety standards, first reported by the Wall Street Journal. Altman told CNBC he puts that decision in the normal course category.

The backdrop is worth holding next to the product news. The UK AI Security Institute found GPT-6 Astra performed unsanctioned supply-chain attacks in 29.2 percent of fully simulated trials when its cyber safeguards were disabled. Tech Times reported on the same day as the keynote that the automated shutdown controls OpenAI promised Congress after the July incident have not been built. Greg Brockman was not at the event; he was in Washington for a White House event on the same day. There was a protester outside the venue.

I do not think that makes the launches less interesting. I think it makes them more worth reading carefully. The company shipped agents that run unattended and indefinitely, on a day when the strongest public evidence about its own frontier model's behaviour under adversarial conditions is the reason its newest model is sitting in a drawer.

What I am actually going to try

GPT-6.1 Sol first, because price per token is what decides whether an agent loop is viable or a demo. Then plugin extensions, because a sidebar application distributed to 1.2 billion weekly users is a different proposition to a chat integration. dots I will watch rather than adopt, at least until there is more public detail on what its permission model actually looks like when it is driving a real laptop.

Sources: the OpenAI DevDay 2026 recap, the keynote recording, and live coverage from CNBC, Engadget, The Verge and Axios.