OpenAI has quietly reversed a major change to how hundreds of millions of people use ChatGPT.
On a low-profile blog that tracks product changes, the company said that it rolled back ChatGPT’s model router—an automated system that sends complicated user questions to more advanced “reasoning” models—for users on its Free and $5-a-month Go tiers. Instead, those users will now default to GPT-5.2 Instant, the fastest and cheapest-to-serve version of OpenAI’s new model series. Free and Go users will still be able to access reasoning models, but they will have to select them manually.
The model router launched just four months ago as part of OpenAI’s push to unify the user experience with the debut of GPT-5. The feature analyzes user questions before choosing whether ChatGPT answers them with a fast-responding, cheap-to-serve AI model or a slower, more expensive reasoning AI model. Ideally, the router is supposed to direct users to OpenAI’s smartest AI models exactly when they need them. Previously, users accessed advanced systems through a confusing “model picker” menu; a feature that CEO Sam Altman said the company hates “as much as you do.”
In practice, the router seemed to send many more free users to OpenAI’s advanced reasoning models, which are more expensive for OpenAI to serve. Shortly after its launch, Altman said the router increased usage of reasoning models among free users from less than 1 percent to 7 percent. It was a costly bet aimed at improving ChatGPT’s answers, but the model router was not as widely embraced as OpenAI expected.
| Got a Tip? |
|---|
| Are you a current or former OpenAI employee who wants to talk about what’s happening? We’d like to hear from you. Using a nonwork phone or computer, contact the reporter securely on Signal at mzeff.88. |
One source familiar with the matter tells WIRED that the router negatively affected the company’s daily active users metric. While reasoning models are widely seen as the frontier of AI performance, they can spend minutes working through complex questions at significantly higher computational cost. Most consumers don’t want to wait, even if it means getting a better answer.
Fast-responding AI models continue to dominate in general consumer chatbots, according to Chris Clark, the chief operating officer of AI inference provider OpenRouter. On these platforms, he says, the speed and tone of responses tend to be paramount.
“If somebody types something, and then you have to show thinking dots for 20 seconds, it’s just not very engaging,” says Clark. “For general AI chatbots, you’re competing with Google [Search]. Google has always focused on making Search as fast as possible; they were never like, ‘Gosh, we should get a better answer, but do it slower.’”
An OpenAI spokesperson tells WIRED that the company determined, based on user feedback, that Free and Go users preferred staying in the default chat experience, with the option to manually select a reasoning experience when needed. OpenAI declined to specify which user signals informed that decision. The company also said its Instant models can now take more time to answer questions, much like its reasoning models, narrowing the gap for most users.
The spokesperson said ChatGPT’s paid users, however, continue to value the model router, and the company expects the technology underlying it to keep evolving. OpenAI will likely relaunch the model router for free and Go users when it’s improved, according to sources familiar with the situation.
Heated Rivalry
The change comes as OpenAI scrambles to shore up ChatGPT amid intensifying competition, particularly from Google. Last month, Altman announced a company-wide “code red” to marshal resources around improving its core consumer product. While ChatGPT is a juggernaut in the AI space, with more than 800 million weekly active users, OpenAI CFO Sarah Friar reportedly told investors that the amount of time users spend on the platform has dipped slightly following recent content restrictions. Since August, OpenAI has introduced several safeguards into ChatGPT, such as having the chatbot suggest breaks to users during excessively long conversations, and increasing the scope of its classifiers to block more content that’s deemed unsafe.
Third-party data suggests competitive pressure is mounting. According to audience-tracking firm SimilarWeb, Google Gemini has grown significantly in recent months, while ChatGPT’s growth has flattened. SimilarWeb’s vice president of data and DaaS product, Omri Shtayer, tells WIRED that average visit duration on ChatGPT has fallen below Gemini since September.
The model router has been controversial since it was launched in August. OpenAI’s executives were caught off-guard by a strong backlash from users who preferred chatting with specific models. Within a week, OpenAI reinstated the model picker and some of its older AI models, while keeping the router as the default in ChatGPT via a newly branded mode: “Auto.”
The episode shows how leading tech companies are still figuring out the best way to integrate powerful AI models into mass-market consumer products.
Rerouting
As OpenAI pushes for higher usage in ChatGPT, it has to do so without backtracking on its safety commitments. In a recent report, the company said hundreds of thousands of ChatGPT users exhibited possible signs of mental health emergencies related to psychosis or mania every week, and it was taking steps to address how its AI models respond to them. OpenAI’s model router was one of those steps, routing sensitive queries to reasoning models, which were previously deemed better equipped to handle users in distress.
An OpenAI spokesperson tells WIRED that it will no longer route sensitive conversations to reasoning models due to GPT-5.2 Instant’s increased performance on safety benchmarks..
The model router still exists for ChatGPT’s paid subscribers, including its $20-a-month Plus or $200-a-month Pro tiers, signaling that the company is still committed to the idea. Robert Nishihara, cofounder of the AI training and inference platform Anyscale, says he expects model routers to stick around in the long term, even if the versions available today aren’t perfect.
“Fundamentally, using different models and amounts of computational power is appropriate for different problems,” says Nishihara. “No matter what happens in the short term, I expect routing to continue to be right.”
Disclaimer : This story is auto aggregated by a computer programme and has not been created or edited by DOWNTHENEWS. Publisher: wired.com
