OpenAI is rolling out fixes for early problems with GPT-5, promising short-term upgrades to capacity, quality, and the user interface.
According to CEO Sam Altman on X, ChatGPT Plus users will see their GPT-5 rate limits doubled once the rollout finishes. Plus subscribers can also keep using the previous 4o model for now. OpenAI will track how much these "legacy" models are used to decide how long they stay available.
Currently, ChatGPT Plus users can send up to 100 messages every three hours. If you hit that cap, the system automatically switches you to a smaller model until the limit resets. GPT-5-Thinking is capped at 200 messages per week, but switching from GPT-5 to GPT-5-Thinking doesn’t count toward that weekly total.
Altman says users should notice an immediate boost in performance, with GPT-5 set to "seem smarter from today." The core of the system is an automatic model switcher that decides which GPT-5 variant handles each prompt. When GPT-5 first launched, problems with this switcher made the model appear "way dumber" for part of the day, Altman claims.
The rollout for all users is taking OpenAI longer than planned. Altman calls it a "massive change at big scale," noting that API traffic has roughly doubled in the past 24 hours. Some bumps were expected with so many changes going live at once, but Altman admits the start was "a little more bumpy" than hoped.
OpenAI adjusts its model routing strategy
The company is also tweaking how it decides which model runs a given prompt, aiming for more reliable matches between task and model. Altman says OpenAI plans to make it clearer which model is in use and to let users manually trigger "Thinking" mode through the interface. He didn’t say when those features will be available.
These updates suggest OpenAI is backing off from its earlier push for full automatic routing. While routing makes things easier for newcomers who don’t want to select models themselves, it can mean less reliable answers or the wrong model for the job—adding extra uncertainty to already unpredictable LLM outputs.
If you want to keep using older ChatGPT models, you can turn on access to "legacy" models in the settings. Once enabled, they'll appear under "Other models." The GPT-5 "Thinking" and "Pro" variants, however, can only be picked directly from a Pro account.