ChatGPT’s model picker is back, and it’s complicated

When OpenAI launched GPT-5 final week, the corporate stated the mannequin would simplify the ChatGPT expertise. OpenAI hoped GPT-5 would act as a kind of “one measurement suits all” AI mannequin with a router that may routinely determine how you can greatest reply person questions. The corporate stated this unified method would remove the necessity for customers to navigate its mannequin picker — an extended, sophisticated menu of AI choices that OpenAI CEO Sam Altman has publicly said he hates.

Nevertheless it seems to be like GPT-5 just isn’t the unified AI mannequin OpenAI hoped it could be.

Altman stated in a post on X Tuesday that the corporate launched new “Auto”, “Quick”, and “Considering” settings for GPT-5 that each one ChatGPT customers can choose from the mannequin picker. The Auto setting appears to work like GPT-5’s mannequin router that OpenAI initially introduced; nevertheless, the corporate can also be giving customers choices to circumnavigate it, permitting them to entry quick and gradual responding AI fashions instantly.

Alongside GPT-5’s new modes, Altman stated that paid customers can as soon as once more entry a number of legacy AI fashions — together with GPT-4o, GPT-4.1, and o3 — which have been deprecated simply final week. GPT-4o is now within the mannequin picker by default, whereas different AI fashions might be added from ChatGPT’s settings.

“We’re engaged on an replace to GPT-5’s character which ought to really feel hotter than the present character however not as annoying (to most customers) as GPT-4o,” Altman wrote within the put up on X. “Nonetheless, one studying for us from the previous few days is we actually simply have to get to a world with extra per-user customization of mannequin character.”

ChatGPT’s mannequin picker now appears to be as sophisticated as ever, suggesting that GPT-5’s mannequin router has not universally happy customers as the corporate hoped. The expectations for GPT-5 have been sky excessive, with many hoping that OpenAI would push the bounds of AI fashions prefer it had with the launch of GPT-4. Nonetheless, GPT-5’s rollout has been rougher than anticipated.

The deprecation of GPT-4o and different AI fashions in ChatGPT sparked a backlash among users who had grown hooked up to the AI fashions’ responses and personalities in ways in which OpenAI had not anticipated. Sooner or later, Altman says the corporate will give customers loads of advance discover if it ever deprecates GPT-4o.

GPT-5’s mannequin router additionally seemed to be largely damaged on launch day. That precipitated some customers to really feel the AI mannequin wasn’t as performant as earlier OpenAI fashions, and compelled Altman to deal with the issue in an AMA session on Reddit. Nonetheless, plainly GPT-5’s router should still not be satisfying for all customers.

“We’re not at all times going to get every part on attempt #1 however I’m very pleased with how rapidly the crew can iterate,” wrote OpenAI’s VP of ChatGPT, Nick Turley, in a post on X Tuesday.

Routing prompts to the correct AI mannequin is a troublesome job that requires aligning an AI mannequin to a person’s preferences, in addition to the particular query they’re asking. The router then has to decide on which AI mannequin to ship the immediate to in only a break up second — that approach, if a immediate goes to a quick responding AI mannequin, the response can nonetheless be quick.

Extra broadly, some folks exhibit preferences for AI fashions that transcend quick or gradual responses. Some customers might just like the verbosity of 1 AI mannequin, whereas others would possibly admire the contrarian solutions of one other.

Human attachment to sure AI fashions is a comparatively new idea that isn’t properly understood. For instance, a whole bunch of individuals in San Francisco just lately held a funeral for Anthropic’s AI mannequin, Claude 3.5 Sonnet, when it was taken offline. In different instances, AI chatbots appear to be contributing to mentally unstable folks taking place psychotic rabbit holes.

It appears OpenAI has extra work to do round aligning its AI fashions to particular person person preferences.