Dots get their own computer
The keynote opened at 10 am in San Francisco, 3 am on Wednesday here in Sydney. The headline product was dots.
A dot is an always-on agent that runs on GPT‑6 Astra. Each one gets its own cloud computer and browser, learns from your feedback over time, and can reach more than 4,000 apps through ChatGPT's plugins. You can message it in ChatGPT, Slack or Teams, or call it in ChatGPT, and texting is coming soon. It also messages you when it has progress, a question, or a decision it needs from you.
OpenAI's launch post gives one example from outside the company. An early tester's dot "noticed he'd forgotten to invoice a publication, prepared the invoice, and sent it after his approval."
Dots are rolling out to Pro and Business Premium in eligible markets, and Enterprise, Edu and Healthcare workspaces can turn on a beta. Your first dot is included in the plan. Talking to it doesn't count toward your ChatGPT limits, but work it hands to Codex or ChatGPT Work does.
What a dot can do without asking
This is the section of the launch post I'd read first, because "always on" only works if the limits are clear.
When you aren't working with it, a dot does what OpenAI calls proactive research. It uses the apps you've connected, but only through read-only tools, so it can't send messages, change anything in those apps, or control a browser or computer.
Some actions could affect your accounts or share information. Before a dot takes one, a step called auto-review checks it against your instructions, your Custom Rules and OpenAI's safety requirements. That decides whether the action goes ahead, waits for your approval, or has to be done by you. Custom Rules let you allow, require approval for, or block specific actions. Some tasks, such as changing a password, always stay with you.
A few more details from the launch post:
- Dots sign in to websites with saved passwords the model never sees.
- Your own computer stays separate unless you connect it.
- An Activity View shows background work as it happens.
- A monitoring system can pause or stop a dot if it detects a safety concern.
GPT‑6.1 Sol costs a fifth of Astra
For developers, the model news was GPT‑6.1 Sol. OpenAI says it nearly matches GPT‑6 Astra on agentic coding, computer use and professional work, at a fifth of Astra's standard token prices.
In the API it's gpt-6.1-sol, at $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. The cached price is 95% below the standard input price and half what GPT‑6 Sol charges for cached input. That matters most for agents, which resend the same long context on every step.
OpenAI's own benchmark numbers are the clearest way to see the trade. On Terminal-Bench Science at maximum effort, a task costs $5.47 on average with GPT‑6.1 Sol, $23.21 with Claude Opus 5.5 and $23.80 with GPT‑6 Astra. Astra still scores highest there, at 68.1%, and OpenAI recommends it for the hardest research tasks.
On DeepSWE v1.1, a benchmark of long software-engineering tasks, OpenAI reports that GPT‑6.1 Sol matches Astra's score at about a fifth of the cost. It's available in ChatGPT Work and Codex on Plus and above, but not yet in regular ChatGPT chat.
Ultrafast is the paid speed tier
Ultrafast is a premium service tier for work where speed matters more than price. OpenAI quotes up to 8× faster token generation in Codex (300 tokens a second) and up to 6× in the API.
GPT‑6 Astra Ultrafast is available now in the API, and in ChatGPT Work and Codex on the new Pro 500 plan and Enterprise. GPT‑6.1 Sol Ultrafast is coming soon. In the API it's one field, service_tier: "ultrafast". The docs recommend a WebSocket connection for agents that make many tool calls in a row.
One detail for anyone outside the US: Ultrafast supports US data residency and global processing only, with no EU or other regional processing endpoints.
Codex and the API
| Announcement | What it does |
|---|---|
| Codex in the cloud | Run Codex on your computer, from your phone, or in the cloud, with reusable team environments |
| Codex CLI refresh | Start and steer tasks by voice, and track several tasks at once in a new /agents view |
| Code Review | Review diffs in the ChatGPT desktop app; Codex can take a first pass in the cloud while you're away |
| Codex Security Cloud | Scan whole GitHub repositories on demand or on a schedule, and prepare fixes in the cloud |
| Decisions API | Point Luna at a fixed set of questions with fixed answers, to classify, route or pick an agent's next step |
| Agents API computer use | Agents built on the API can now operate software, alongside multi-agent work and context compaction |
| Bedrock Managed Agents | The Agents API's core, built with Amazon to run entirely inside AWS |
The Decisions API is in limited preview, with a broad release planned in the coming days. The others are available now, on the plans OpenAI lists in its recap.
Plugins can now react to events
The other announcement I'd look at closely is MCP Events. Until now, a plugin's MCP server mostly waited for ChatGPT to call one of its tools. MCP Events, a proposed addition to the protocol, lets ChatGPT subscribe to your server and be told when something happens.
A user tells ChatGPT what to watch, such as review comments on one document. ChatGPT subscribes through your server with a callback URL and a signing secret. Your server checks the callback with a one-time challenge, then posts signed events to it. ChatGPT then follows the user's instructions for each event, even while they're away.
It needs MCP 2.0 (protocol version 2026-07-28) and supports webhook delivery only. The MCP server behind this site only answers when it's asked, so events are the half it doesn't have yet.
Alongside events, plugin extensions let a plugin add its own sidebar app, a panel beside the conversation, or a viewer for a file type. OpenAI describes it as opening up "the platform we use to build ChatGPT features".
The rest of the list
ChatGPT Space is a shared area where teammates, ChatGPT and your dot work from the same knowledge. Pages are documents that people and agents edit together, and collaborative slides are due in the coming weeks. Team tasks run recurring work on a schedule or when a new email or Slack message arrives. You can also mention @ChatGPT in Slack and Teams threads.
Sign in with ChatGPT now lets Plus and Pro users spend their plan's allowance in 16 partner tools, including Vercel, Notion and Cognition's Devin. The new Pro 500 plan has 25 times the Plus allowance and includes Ultrafast. For enterprises, Private Intelligence adds safety reviews that don't give OpenAI staff access to the underlying content.
Where I'd start
Switching an existing agent to gpt-6.1-sol and rerunning its evals is the cheapest change on the list. If you run an MCP server, events are the change that lets it start work instead of waiting to be called. Before trying dots, it's worth reading the Custom Rules and auto-review sections of the launch post.