+ dots: always-on agents, own computer
+ GPT-6.1 Sol: near-Astra, 1/5 the price
+ Ultrafast + a Pro 500 plan
+ Codex cloud, Agents API, Decisions API
- live demos: “a slow morning”You'd think an agent that works around the clock would be awake for its own launch. OpenAI spent an hour introducing one, then asked it a question, live on stage. And at the very end, somebody on stage pressed a real button, and they said it was for everyone in the world. I'll get to it. On Tuesday in San Francisco, OpenAI ran DevDay twenty twenty-six, and the recap page lists more than twenty major launches. Most of them fit in five lines, and one of those lines is the demo you just watched.
Start with dots. A dot is an agent running on GPT-6 Astra, with its own computer in the cloud and plugins into more than four thousand apps. You reach it in ChatGPT, Slack or Teams, and you can call it. Texting is coming soon, so your agent will be able to text you at three in the morning, like an ex. OpenAI's launch film shows what that feels like. The user renames theirs on day one. I'm going to call you Alfred. Alfred then reports that the cake vendor for the wedding cancelled, and that he has already found a backup, which is more than my calendar app has ever done for me. OpenAI also says its own engineers have their dots fixing dozens of bugs a day. Holly, on stage, added that dots were built by dots, which would explain the slow morning. Within an hour, Hacker News had two open source clones of it, the developer version of a standing ovation. Your first dot comes with the Pro and Business Premium plans in eligible markets, and talking to it doesn't count against your usage limits.
For companies there are specialist dots, with their own identity and credentials, managed through Microsoft's Agent 365. Yesterday I told you the Dutch government is building its own Linux to get away from Microsoft. Today OpenAI's agents got a Microsoft manager. Then the office suite. ChatGPT Space is a shared home for your team, their dots and ChatGPT, built on Pages, a new kind of document for humans and agents together. Collaborative slides follow in the coming weeks, with export to PowerPoint and Google Slides, so Notion and Google Docs have been served notice, in a document your agent drafted. Now the model. GPT-6.1 Sol costs two dollars per million tokens in and ten dollars out. OpenAI says that buys nearly Astra's intelligence at a fifth of Astra's price. Cached input dropped to ten cents, half of what the old Sol charged, and one Hacker News commenter called that the actual big announcement. It's in the API, Codex and ChatGPT Work, and plain ChatGPT chat is still waiting for it.
A week ago I told you Claude Opus 5.5 cut its prices, and OpenAI answered with GPT-6 Sol at two dollars. That Sol was the new one for exactly seven days, and its replacement costs the exact same two dollars. Hacker News asked if it's the shortest model life ever. Every chart on that page is OpenAI grading OpenAI, and one footnote even admits the number for Anthropic's Fable leaves out the cost of fallbacks, which happened on about forty percent of tasks. So here's someone else's scoreboard. On Artificial Analysis, Sol scores fifty-two and Astra fifty-three. Sol's cost per task is about a fifth of Astra's, so the claim holds up. The same board still puts Claude Opus 5.5 on top, at fifty-eight. Then speed. Ultrafast runs Astra at up to three hundred tokens a second. OpenAI's pricing page lists it at six times the normal rate, which is three hundred dollars per million output tokens. On stage, the whole pitch for that price was one line. You know what? It's worth it. The demo was two rockets racing, and the expensive one won.
Ultrafast in ChatGPT needs the new top plan, Pro 500, which OpenAI says gets twenty-five times the Plus allowance. And the two-hundred-dollar Pro plan reopens today. Three weeks ago I told you OpenAI paused new Pro sign-ups because everyone wanted Astra. The waiting room is closed, and there's a bigger room upstairs. For developers, the sleeper hit is the Decisions API. You hand the small Luna model a fixed list of answers, it picks one in a fraction of a second, and you use that to route requests, classify content or choose an agent's next step. On stage they promised the video was real time, which is exactly what people say when it looks sped up. It's in limited preview today. Then Sam Altman took a victory lap. Last year OpenAI predicted an AI research intern within a year, and he says they now have one. A research lead added that since July, their models complete over a third of day-long research tasks with no help. There's no paper yet, so for now the intern exists according to its employer, and like every intern, it's about to get blamed for everything.
Codex got the most launches. It now runs in the cloud with reusable environments, a Code Review view takes a first pass on your pull requests, and Codex Security Cloud scans whole GitHub repos on a schedule. The command line now takes voice commands, so naturally, the voice demo went like this. Romain, the presenter, typed the prompt instead, and later gave the cloud version the one task every backend team dreams about. Afterwards, Tibo Sottiaux, who runs Codex, wrote that the live demos suffered from rolling out all the updates at the same time. Under all of it sits the Agents API, the same harness that runs dots, now with computer use, so your own agent can click through a browser while OpenAI runs the servers. Amazon gets its own version, Bedrock Managed Agents, which runs OpenAI agents entirely inside AWS. And for nervous enterprises, Private Intelligence adds zero data retention, with safety reviews that no OpenAI employee gets to read.
Now the one that touches your wallet. With Sign in with ChatGPT, a user logs into your app and spends their own ChatGPT plan inside it, so you stop paying for their tokens. Sixteen partners are live, including Devin, Notion and Vercel, and the user sets how much each app may use. One Hacker News commenter called it how this was always supposed to be. Then distribution. Plugin extensions give your app a home in ChatGPT's sidebar with its own panels, ChatGPT will now suggest your plugin in the middle of a conversation, and MCP events let a plugin wake up when something happens elsewhere. That's a store with one point two billion weekly users, and a single landlord. And to show the API driving hardware, Romain brought a robot duck named Lavender.
Now, that button. On stage, OpenAI joked it has done so many resets that there's a campaign to rename it the reset company. Then Tibo, whose usage limit resets are a running joke among Codex users, pressed a real button under a flip cover. The question on stage was, does this just work like that? They said it was for everyone in the world, which is how a keynote ends with a crowd of developers refreshing their usage page. Reviews were split. A Hacker News commenter called it an embarrassment, and one reply on X called the reset the only positive of the day. Then Sam Altman closed. He said he finds the industrial revolution comparison off-putting, and that AI done right can be much more like a new Renaissance. A Renaissance, with usage limits and a reset button. Your Monday list. Move batch and agent jobs to Sol and lean on the cheap cache, try the Decisions API for routing, and look at Sign in with ChatGPT before your competitor does.
Verdict: NEEDS REVIEW — prices check out; the intern and the demos are OpenAI grading OpenAI
Sources
https://www.youtube.com/watch?v=Fls_onRviPM
https://news.ycombinator.com/item?id=49896309
https://openai.com/index/devday-2026-recap/
https://news.ycombinator.com/item?id=49896600
https://openai.com/index/introducing-gpt-6-1-sol/
https://news.ycombinator.com/item?id=49896586
https://openai.com/index/introducing-dots/
https://openai.com/index/how-we-build-safety-security-and-privacy-into-dots/
https://x.com/thsottiaux/status/2104989161774322009
https://developers.openai.com/api/docs/pricing
https://artificialanalysis.ai/models/gpt-6-1-sol
https://artificialanalysis.ai/models/gpt-6-astra
https://artificialanalysis.ai/models/claude-opus-5-5
https://news.ycombinator.com/item?id=49897468
https://news.ycombinator.com/item?id=49897710
https://news.ycombinator.com/item?id=49897591
https://help.openai.com/en/articles/20001542-using-your-chatgpt-plan-in-other-apps-and-sites
https://x.com/thsottiaux/status/2104994835212226681
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · thedailydiff.dev · forward this to the intern who deployed on Friday.