Overall takes
The key words of the day were “Codex” and “partners.”
The Codex desktop harness has been great for a while, but now they’ve extended what’s been great in the desktop app to cloud agents, custom agents, always-on bots, etc. And a lot of the partnership discussion was not just about GPT’s ecosystem of connectors and plugins, but also ways other people’s apps can use your or your company’s ChatGPT account (and limits).
What shipped? A lot:
- Dots: OAI’s answer to Muse or Grok Bot, aimed more at work than personal stuff.
- ChatGPT Space: A Notion-type document workspace where agents/dots and multiple users can jam on docs, slides, spreadsheets. Similar to Claude’s new docs that shipped last week, but with a clearer, simpler setup that’s more like a wiki. And, unlike Claude which generates these as an artifact, I think the idea is you can create a collaborative Space and populate it yourself as the starting point or context for agents or dots to do stuff.
- GPT-6.1 Sol: The model releases will continue until morale improves! Capabilities at/above Astra on certain tasks at 1/5th the cost.
- Private Intelligence: ZDR with private safety processing, and (soon) “Private Inference” with confidential computing and data handling controls
- Codex Cloud is out of beta with the same kinds of reusable dev environments as Claude Code Web, Cursor Cloud, Amp, Devin, etc. (This may not be rolled out full yet, as cloud sessions in my Codex app are still weird and janky.)
- New updates to the GUI and TUI Codex apps: The Codex terminal app is a lot nicer now! Some of the GUI changes dropped over the weekend, the main thing is adding a top tab bar so you can more easily see and page between open sessions.
- New Code Review view in Codex, which shows all your open PRs so you can request and monitor Codex-driven reviews in one place
- Codex Security Cloud which can do automated/scheduled scans against your repos and autonomously create PRs with fixes
- Agents API beta: Technically this came out a few weeks ago, but it’s OAI’s version of Claude Managed Agents, with (afaik) a much simpler, more friendly API. These custom agents run on the Codex harness, have sandboxes — including computer use — on OAI’s servers, and can also be run on AWS or other cloud/sandbox providers
- Decisions API is OpenAI’s answer to Jev — a super-fast, cheap, focused way to have a model look at some data and choose from a list of options. Whereas Jev is its own kind of model, Decisions runs on GPT-6 Luna (which means it can handle image input, which Jev can’t).
- Plugin extensions go beyond the old “just connectors and skills in a package” plugin model and put custom UI in the ChatGPT client. For example, the Figma plugin now has a UI allowing you to load a Figma file and have ChatGPT do stuff to it.
- The Pro 200 plan is back: Limits are now just 10x, not 20x, but no more 5 hour limits, and Dots agent usage (which uses Astra) isn’t counted against your plan unless it has to spin up a separate ChatGPT or Codex worker.
- There’s also a Pro 500 plan, with new ULTRAFAST speed: Exactly what it sounds like — about 25x Plus limits, and only this plan can use the new mode that returns results 6x faster (with commensurate token burn).
@ChatGPT in Slack and Teams — again, like Claude Tag, but with ChatGPT.- Partners can use your ChatGPT subscription limits — you’ll be able to sign into (e.g.) Devin, Notion, Vercel, Lovable, others with your ChatGPT account and those products will use your subscription limits rather than charge you separately for inference. (There are some existing products on the list, like Amp or OpenCode, that already support ChatGPT subscriptions.)