OpenAI backtracks on GPT-5, restores model selection & customization
OpenAI has recently navigated a period of intense user feedback, initiating a significant course correction following the rollout of its latest flagship model, GPT-5. The initial strategy for GPT-5 aimed for a unified artificial intelligence experience, where a single, advanced model would intelligently route user requests to the most appropriate AI backend, ideally simplifying interactions by eliminating the need for manual model selection. However, this vision quickly encountered resistance from a user base accustomed to greater control.
The debut of GPT-5 sparked widespread user protests, particularly concerning the removal of options to select previous, preferred models. The immediate backlash prompted OpenAI CEO Sam Altman to acknowledge the issues, confirming the swift return of the GPT-4o model. This initial concession was soon followed by more extensive adjustments, signaling a notable departure from the unified model philosophy.
Responding to the clamor for more granular control, OpenAI has now introduced a “show additional models” switch within ChatGPT web settings for paid subscribers. This new toggle provides access to a range of older models, including o3, 4.1, and a specialized GPT-5 Thinking mini. Furthermore, GPT-5 itself now offers users a choice between ‘Auto’, ‘Fast’, and ‘Thinking’ modes. While ‘Auto’ is recommended for most general use cases, the ‘Fast’ and ‘Thinking’ options cater to users seeking more specific performance characteristics or deeper computational processing, with the ‘Thinking’ mode currently capped at 3,000 messages per week. Altman reiterated that GPT-4o remains the default for all paid users and assured that ample notice would be provided should it ever be phased out.
These rapid updates represent a clear admission from OpenAI that its initial approach to GPT-5’s request routing did not align with the diverse preferences of its user base. Nick Turley, OpenAI’s VP and Head of ChatGPT, candidly acknowledged the company’s misstep, stating, “We’re not always going to get everything on try #1,” while simultaneously expressing pride in the team’s ability to iterate quickly based on feedback. Altman himself previously admitted that suddenly discontinuing models users relied upon was a mistake.
Despite the internal adjustments and user-driven changes, the adoption of GPT-5 continues across the industry. Microsoft, for instance, recently confirmed the model’s integration into GitHub Copilot for Visual Studio, hailing it as OpenAI’s most advanced model to date. Looking ahead, OpenAI is also working on refining GPT-5’s “personality” to be “warmer” than its current iteration, yet less prone to the perceived annoyances some users experienced with GPT-4o. This ongoing refinement underscores a key learning for the company: the imperative for greater per-user customization of AI model personalities to meet the varied demands of its global audience.