In practice, though, OpenAI launched three different modes for GPT-5 in the model picker: Auto, Fast, and Thinking. The Auto mode behaves like the envisioned router, choosing between GPT-5's sub-models in real-time. The Fast mode is optimized for quick, lightweight responses, while the Thinking mode uses more heavyweight inference engines that take a little more time to give more in-depth responses.
Updates to ChatGPT:
— Sam Altman (@sama) August 13, 2025
You can now choose between “Auto”, “Fast”, and “Thinking” for GPT-5. Most users will want Auto, but the additional control will be useful for some people.
Rate limits are now 3,000 messages/week with GPT-5 Thinking, and then extra capacity on GPT-5 Thinking…
At the same time, paying users are given restored access to a number of deprecated legacy models—specifically GPT-4o, GPT-4.1, and o3—that were deprecated just days earlier. GPT-4o returns as the default model, while other deprecated models can be enabled through ChatGPT settings. The about-face is an example of OpenAI's willingness to move quickly to meet user demand for familiar personalities and performance characteristics.
Despite these improvements, ChatGPT's model selection interface is still as complicated as ever. The early promise of an easy, router-supported "Auto" experience has been replaced with a vast set of selectable modes. Users must now decide not just whether to yield control to the router but also which older model they would choose, thus recreating the very complexity that GPT-5 was designed to remove. The notice set off an uproar among users who had become fond of certain AI personas. They had learned to work with the eccentricities of GPT-4o or the profundity of o3. Taking away those choices angered users—and OpenAI responded, bringing back the old friends sooner than intended. The incident highlights how intensely users form connections with AI avatars, and in unforeseen manners.
Contributing to the chaos, GPT-5's internal router had hiccups on launch day. A few users noted slower responses or replies that seemed off from expectations. OpenAI's VP of ChatGPT conceded the bumpy rollout but said he was proud of the team's fast iteration. Routing prompts to the best sub-model in milliseconds is still an engineering hurdle.
Going forward, OpenAI plans to increase per-user customization of model "personalities." The company is tuning GPT-5's default tone to find a middle ground between verbosity and warmth, and it wants to enable each user to customize the AI's personality. The key question to the ChatGPT ecosystem is still whether this shift will actually reduce decisions or just trade one choice set for another.
Ultimately, the reintroduction of a comprehensive model selector in GPT-5 illustrates that the essence of simplicity in artificial intelligence encompasses not only technological aspects but also psychological considerations. Users seek a combination of capability and familiarity, and addressing these requirements necessitates adaptability. As OpenAI continues to refine GPT-5, the equilibrium between effortless automation and manual intervention will shape the trajectory of conversational artificial intelligence.