OpenAI's GPT-5 ships as a router over multiple models, not one model

GPT-5 is a system of models with a router deciding between them

The most concrete fact in this system card is that GPT-5 is not one model. OpenAI describes it as a unified system containing a fast model (gpt-5-main), a deeper reasoning model (gpt-5-thinking), and a real-time router that decides which to invoke based on 'conversation type, complexity, tool needs, and explicit intent.'

The router is not static. OpenAI says it is 'continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness.' That means the routing behavior a team observes in ChatGPT today can shift underneath them as the router learns from aggregate usage.

There is also a fallback tier: once usage limits are hit, 'a mini version of each model handles remaining queries.' So the same prompt can be served by a smaller model depending on load, not just complexity.

ChatGPT and the API expose different pieces of the family

For teams building on the platform, the split between the consumer product and the developer surface matters. In ChatGPT, users get the router experience plus gpt-5-thinking-pro, which OpenAI says 'makes use of parallel test time compute.' The API, by contrast, gives 'direct access to the thinking model, its mini version, and an even smaller and faster nano version.'

OpenAI maps each new model to a predecessor: gpt-5-main succeeds GPT-4o, gpt-5-thinking succeeds o3, gpt-5-thinking-mini succeeds o4-mini, and gpt-5-thinking-nano succeeds GPT-4.1-nano. That mapping is the practical migration guide — it tells developers which current model each GPT-5 tier is meant to replace.

The company states its intent to 'integrate these capabilities into a single model' in the near future. Until then, the routing and tiering are an architectural reality developers have to reason about, not an implementation detail hidden behind one endpoint.

A precautionary High capability call on biology and chemistry

OpenAI has classified gpt-5-thinking as 'High capability in the Biological and Chemical domain' under its Preparedness Framework and activated the associated safeguards. Notably, it did this without conclusive evidence.

While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm—our defined threshold for High capability—we have chosen to take a precautionary approach.Montana Labs

This mirrors the treatment of ChatGPT agent. The framing tells developers that safeguards on the thinking model are triggered by a precautionary standard rather than demonstrated harm, which affects what behaviors and refusals to expect in bio/chem-adjacent applications.

What the router architecture requires teams to test

The specific implication of this announcement is that evaluating 'GPT-5' means evaluating a moving target. Because the router is continuously retrained and falls back to mini models under load, a prompt's behavior depends on signals outside a developer's control. Teams that need deterministic model selection should use the API's named endpoints — gpt-5-thinking, its mini, and its nano — rather than relying on the routed ChatGPT experience, and should test against the specific tier they intend to ship on.