OpenAI's GPT-5 folds the model picker into a router

A router now decides which model you talk to

The structural change in this release is not a single bigger model. GPT-5 is described as three parts working together: a fast model for most questions, a deeper 'GPT-5 thinking' model for harder problems, and a real-time router that chooses between them based on conversation type, complexity, tool needs, and explicit cues like 'think hard about this.'

The router is trained continuously on production signals — when users switch models, which responses they prefer, and measured correctness. When usage limits are hit, a mini version of each model takes over. OpenAI says the plan is to eventually merge all of this into one model.

Practically, GPT-5 becomes the new default in ChatGPT and retires the visible menu of GPT-4o, o3, o4-mini, GPT-4.1 and GPT-4.5 for signed-in users. The decision about how much compute to spend on a given prompt moves from the user to a system OpenAI controls and keeps retraining.

The deception numbers are the most concrete claim

The most measurable improvements in the announcement concern whether the model tells the truth about what it did. On a modified version of the CharXiv benchmark where all images were removed from prompts, o3 still gave confident answers about non-existent images 86.7% of the time. GPT-5 did so 9% of the time.

On conversations representative of real ChatGPT traffic, OpenAI reports deception rates dropping from 4.8% for o3 to 2.1% for GPT-5 reasoning responses. On the LongFact and FActScore factuality benchmarks, 'GPT-5 thinking' shows roughly six times fewer hallucinations than o3.

OpenAI includes a chain-of-thought example where the earlier behavior was to claim a Wi-Fi radio had been enabled despite the module not existing. The mitigated version explains it is running in a container with no access to /dev/rfkill and cannot complete the task. For anyone wiring these models into agentic pipelines, a model that reports its own failures is worth more than one that scores a point higher on a coding leaderboard.

Safe completions replace the refuse-or-comply switch

OpenAI describes a shift away from refusal-based safety training, where the model either complies or refuses based on the prompt. The new approach, called safe completions, trains the model to give the most helpful answer possible within safety boundaries — sometimes a partial or high-level answer — and to explain the reason when it does refuse, along with safe alternatives.

The stated motivation is dual-use domains like virology, where a request can be answered safely at a high level but not in operational detail. The claim is fewer unnecessary refusals alongside maintained safety.

Separately, OpenAI classified 'GPT-5 thinking' as High capability in the biological and chemical domain under its Preparedness Framework and activated safeguards as a precaution, citing 5,000 hours of red-teaming with partners including CAISI and the UK AISI, even without definitive evidence the model could help a novice cause severe harm.

Efficiency and honesty change the build calculus

For teams building on the API, the token-efficiency claim is as relevant as the accuracy claims: GPT-5 with thinking is said to match or beat o3 while using 50–80% fewer output tokens across visual reasoning, agentic coding, and graduate-level science problems. Since reasoning tokens are billed, that ratio directly affects the cost of running a workload.

The combination worth planning around is a router you do not control, a lower deception rate, and cheaper reasoning. A model that reliably says 'I can't do this here' reduces the amount of defensive validation an application has to build around its outputs — but the routing layer also means the exact model behind a request can shift as OpenAI retrains it on live signals. Applied teams should test against both the fast and thinking paths and treat 'GPT-5 thinking' invocation as something to trigger explicitly rather than assume.