Chatgpt’s Model Picker Is Back, And It’s Complicated

When OpenAI launched GPT-5 past week, nan institution said nan exemplary would simplify nan ChatGPT experience. OpenAI hoped GPT-5 would enactment arsenic a benignant of “one size fits all” AI exemplary pinch a router that would automatically determine really to champion reply personification questions. The institution said this unified attack would eliminate nan request for users to navigate its exemplary picker — a long, analyzable database of AI models that OpenAI CEO Sam Altman has said he hates — to prime a type of ChatGPT that offers nan correct benignant of responses.

But it looks for illustration GPT-5 is not nan unified AI exemplary OpenAI hoped it would be.

Altman said successful a post connected X Tuesday that nan institution introduced caller “Auto”, “Fast”, and “Thinking” settings for GPT-5 that each ChatGPT users tin prime from nan exemplary picker. The Auto mounting seems to activity for illustration GPT-5’s exemplary router that OpenAI initially announced; however, nan institution is besides giving users options to circumnavigate it, allowing them to entree accelerated and slow responding AI models directly.

Alongside GPT-5’s caller modes, Altman said that paid users tin erstwhile again entree respective bequest AI models — including GPT-4o, GPT-4.1, and o3 — which were deprecated conscionable past week.

“We are moving connected an update to GPT-5’s characteristic which should consciousness warmer than nan existent characteristic but not arsenic annoying (to astir users) arsenic GPT-4o,” Altman wrote successful nan station connected X. “However, 1 learning for america from nan past fewer days is we really conscionable request to get to a world pinch much per-user customization of exemplary personality.”

ChatGPT’s exemplary picker now seems to beryllium arsenic analyzable arsenic ever, suggesting that GPT-5’s exemplary router has not universally satisfied users arsenic nan institution hoped. The expectations for GPT-5 were entity high, pinch galore hoping that OpenAI would push nan limits of AI models for illustration it had pinch nan motorboat of GPT-4. However, GPT-5’s rollout has been rougher than expected.

The deprecation of GPT-4o and different AI models successful ChatGPT sparked a backlash among users who had grown attached to nan AI models’ responses and personalities successful ways that OpenAI had not anticipated. In nan future, Altman says nan institution will springiness users plentifulness of beforehand announcement if it ever deprecates GPT-4o.

ChatGPT’s exemplary picker erstwhile again has respective options (Credit: openai/maxwell zeff)

GPT-5’s exemplary router besides appeared to beryllium mostly surgery connected motorboat day. That caused immoderate users to consciousness nan AI exemplary wasn’t arsenic performant arsenic erstwhile OpenAI models, and forced Altman to reside nan problem successful an AMA convention connected Reddit. However, it seems that GPT-5’s router whitethorn still not beryllium satisfying for each users.

“We’re not ever going to get everything connected effort #1 but I americium very proud of really quickly nan squad tin iterate,” wrote OpenAI’s VP of ChatGPT, Nick Turley, successful a post connected X Tuesday.

Routing prompts to nan correct AI exemplary is simply a difficult task that requires aligning an AI exemplary to a user’s preferences, arsenic good arsenic nan circumstantial mobility they’re asking. The router past has to make a determination connected which AI exemplary to nonstop nan punctual to successful conscionable a divided 2nd — that way, if a punctual goes to a accelerated responding AI model, nan consequence tin still beryllium fast.

More broadly, immoderate group grounds preferences for AI models that spell beyond accelerated aliases slow responses. Some users whitethorn for illustration nan verbosity of 1 AI model, while others mightiness admit nan contrarian answers of another.

Human attachment to definite AI models is simply a comparatively caller conception that isn’t good understood. For example, hundreds of group successful San Francisco precocious held a funeral for Anthropic’s AI model, Claude 3.5 Sonnet, erstwhile it was taken offline. In different cases, AI chatbots look to beryllium contributing to mentally unstable group going down psychotic rabbit holes.

It seems OpenAI has much activity to do astir aligning its AI models to individual personification preferences.