OpenAI Promises 28 Days of ChatGPT and Codex Upgrades
OpenAI's head of ChatGPT and Codex says his team will ship one clear improvement every day for four weeks, or reset everyone's usage on a day it can't. Day one made the models about 50% faster.
Tibo Sottiaux, who runs ChatGPT and Codex at OpenAI, made a public promise on Sunday, a few days after DevDay. For the next 28 days, his team will ship one thing a day that's a clear improvement for most Codex and ChatGPT Work users.
On any day it can't do that, it ships "a full reset" instead.
Day one landed Monday, and it was a speed boost.
What changed on day one
GPT-6 Astra and GPT-6.1 Sol, the main models behind ChatGPT Work and Codex (OpenAI's coding agent), now answer about 50% faster at their default speed. Sottiaux said it applies to everything you run through your ChatGPT subscription, there's nothing to switch on, and people should feel it within two hours of his post.
In practice, Work and Codex tasks should come back noticeably sooner. Work comes with every paid ChatGPT plan except Go, so if you're on Free or Go, none of this reaches you yet.
The boost also follows your subscription into other apps. A growing list of tools let you sign in with your ChatGPT account and run on your plan instead of a separate bill, and Sottiaux named several that get the faster models: the coding agent Devin, plus OpenCode, Pi and Amp. We covered how that sign-in works in our DevDay explainer.
Monday's change is separate from Ultrafast, the much quicker Astra mode OpenAI showed at DevDay. According to OpenAI's help page for Work and Codex, Ultrafast is only on the new $500 Pro plan and some Enterprise and Edu workspaces. What got faster on Monday is the regular speed everyone else uses.
What "a full reset" means
Sottiaux's post only says "full reset." People who use Codex read it as a reset of usage limits, the cap on how much Work and Codex you can use before you have to wait for your allowance to refill. A reset puts the whole allowance back early.
Resets are already part of how his team works. On Lenny's Podcast, recorded on DevDay, Sottiaux said he can press "the reset button" whenever it's needed. Asked how many resets people should expect, he said, "Depends how many times we break things."
So each of the 28 days ends with subscribers getting something: an upgrade, or a refilled allowance.
Chat and Work are becoming one ChatGPT
In the same interview, Sottiaux said OpenAI will merge Chat and Work. Right now ChatGPT has a toggle at the top of the page: Chat for quick answers, Work for longer jobs where an agent takes action across your apps and files. (Our first look at ChatGPT Work covers what it does.)
"We heard the feedback loud and clear," he said. People like what Work can do, but they reach for Chat because it's faster and more pleasant to use. "We're merging that."
He also wants to get rid of the model picker, the menu where you choose a model and how hard it thinks. "You kind of need a PhD in model pickers," he said, adding that he gets tired of it himself.
OpenAI's new Dots agents already work that way. Sottiaux says dots have no model picker, and the only thing you configure is where you talk to it. (I gave my dot real jobs for a day last week.) Sottiaux said OpenAI will eventually ship everything Dots can do straight into ChatGPT. He gave no date for any of it.
Why OpenAI is doing this now
OpenAI had a bumpy launch week. Along with Dots and the new models, it added a $500 Pro plan and cut what the $200 plan gets, which upset a lot of its heaviest users.
Speed was the other complaint, with a lot of Codex users saying Sol felt slow. After day one, a lot of Codex users timed the models before and after, and many say they're still about half as fast as Claude.
Both complaints sting more because of Claude. Anthropic launched Claude Opus 5.5 two weeks ago, and it has pulled a lot of attention from the same people who pay for Codex. I put Opus 5.5, Fable 5.1 and GPT-6 Astra on the same three builds right after launch:
A promise like this turns every day into a public scorecard, and Codex users are already grading each ship with a stopwatch. That pressure is good for anyone paying for ChatGPT.
I like that the fallback costs OpenAI something real. If OpenAI misses a day, it hands back usage, which is exactly what $200 subscribers are about to lose on October 30.
Comments
Sign in to join the conversation.