OpenAI is giving ChatGPT users more control over which GPT-5 models they use. Three manual modes, new model selection options, and the return of GPT-4o mark a shift away from the automatic routing system that came with GPT-5.
ChatGPT now lets users manually toggle GPT-5 between "Auto," "Fast," and "Thinking" modes. CEO Sam Altman says "most users will want Auto," but the manual options give more flexibility "for some people."
There's also a small UI tweak: if you hover over the Regenerate button, ChatGPT now shows exactly which model generated the answer.
A return to user control
This marks a shift away from the routing system introduced with GPT-5, where ChatGPT would automatically choose which model to use for each prompt.
The change follows a wave of criticism. Some users found automatic routing opaque and unreliable, especially when proven models like GPT-4o disappeared without warning and the router didn't always pick the most capable GPT-5 version for the task.
Some also suspected that OpenAI used routing to steer costly requests toward cheaper models and limit how often users could access the full "reasoning" variant. Initially, paying customers saw their quota for reasoning requests sharply reduced.
In response to criticism, OpenAI later raised those limits. Users can now send up to 3,000 messages per week in "GPT-5 Thinking" mode before ChatGPT automatically switches them to the smaller "GPT-5 Thinking mini" model.
OpenAI notes on its website that this is just a "temporary increase" and will "revert to the previous limit in the near future." Altman also says these limits could change again depending on how people use the service.
OpenAI has also brought GPT-4o back to the model picker for all paying users. If the company decides to phase it out again, they say they'll provide notice.
There's a new "Show additional models" button in the ChatGPT web settings for paid accounts as well. Turning it on lets users pick from legacy models like o3, 4.1, and GPT-5 Thinking mini. GPT-4.5 remains exclusive to Pro users, which Altman says is because "it costs a lot of GPUs."
Critics see little progress
Despite the updates, core criticisms of GPT-5 persist. In a recent blog post, LLM skeptic Gary Marcus called GPT-5 "overdue, overhyped and underwhelming," arguing that it's more of a rushed incremental update than a real breakthrough.
Marcus points to early errors in physics, chess, and image analysis. A recent study from Arizona State University backs him up: "chain of thought" reasoning—often touted as a strength of large language models—turns out to be fragile and unreliable outside of training data, a flaw Marcus also sees in Grok and Gemini.
For Marcus, the launch of GPT-5 isn't a milestone on the road to general AI (AGI), but a moment when even tech insiders might start doubting the industry's scaling strategy.
Meanwhile, OpenAI and Google DeepMind are using the same technology to solve math and logic problems that seemed unreachable just months ago. The specifics are still under wraps, but the results show today's language models can pull off surprising feats despite their flaws and that the technology still has a lot of unknown territory left to explore.
