GPT-5 isn't one model — it's a router quietly picking your brain

OpenAI’s GPT-5 launch reveals less of a single model and more of an on-the-fly traffic controller. The system unifies a fast, general-purpose model (gpt-5-main) with a deeper reasoning engine (gpt-5-thinking), tied together by a real-time router that decides which brain to use based on your prompt’s complexity, tool needs, and even explicit cues like ‘think hard about this.’ The router doesn’t just follow static rules. OpenAI is continuously training it on live signals — including when users manually switch models, which responses they prefer, and measured correctness scores — so the system theoretically gets smarter about delegation over time. Once you hit usage caps, mini versions of each model step in to handle the remaining workload.

What’s really happening under the hood is a quiet consolidation of OpenAI’s model lineup. The system card positions gpt-5-main as the successor to GPT-4o and gpt-5-thinking as the heir to the o-series reasoning models. For developers, the API exposes the thinking model in three sizes — standard, mini, and an even smaller nano variant clearly aimed at cost-sensitive builds. ChatGPT users get access to a premium tier called gpt-5-thinking-pro, which uses parallel test-time compute to squeeze out better answers.

OpenAI claims the real win here isn’t benchmark bragging rights. The company says it has ‘leveled up’ performance in ChatGPT’s three workhorse categories — writing, coding, and health — while making meaningful strides against hallucinations, sycophancy, and poor instruction following. The safe-completions training technique is baked into every variant to block disallowed content. It’s a practical pitch: less flash, more reliability for the stuff people actually use ChatGPT to do every day.

There’s one eyebrow-raising detail buried in the safety assessment. OpenAI has decided to classify gpt-5-thinking as ‘High capability’ under its Biological and Chemical Preparedness Framework, triggering extra safeguards — even though the company admits it doesn’t have ‘definitive evidence’ the model could meaningfully help a novice create severe biological harm. The threshold for High capability remains unmet by their own definition. They’re calling it a precautionary move, but it’s the kind of admission that makes you wonder what their internal red-teaming actually turned up.