OpenAI: GPT-5 Mini

GPT-5 Mini is a compact variant of GPT-5 designed for lighter-weight reasoning at a fraction of the latency and cost of the full model. It is the direct successor to OpenAI’s o4-mini, and unlike that older model it inherits the GPT-5 generation’s instruction-following, safe-completions behaviour and reduced hallucination profile. In the routed GPT-5 system it is one of the tiers the router selects when a request does not need the deepest reasoning model.

For developers the practical pitch is economics. Mini costs roughly $0.25 per million input tokens and $2.00 per million output tokens — about a fifth of full GPT-5 — while keeping the 400,000-token context window. That makes it the sensible default for the bulk of an application’s traffic, with full GPT-5 reserved for the small fraction of requests that genuinely benefit from extended thinking.

The GPT-5 family is best understood as a cost/latency ladder over a shared capability base. Full GPT-5 (and gpt-5-thinking) sits at the top for the hardest reasoning; GPT-5 Mini occupies the middle, trading some depth for speed and price; GPT-5 Nano sits at the bottom for ultra-low-latency, high-volume tasks. All three share the same safety tuning and the same large context window.

In practice this means you rarely choose Mini in isolation — you design a routing policy. Send classification, extraction, short summarisation and routine chat to Mini; escalate to full GPT-5 only when a task fails Mini’s quality bar or explicitly needs deep multi-step reasoning. This tiering is the single biggest lever for keeping average cost per request low without sacrificing quality on the requests that matter.

Mini shines on tasks that are well-specified and high-volume: structured extraction, routing and triage, content classification, retrieval-augmented question answering over your own documents, and the many small reasoning steps inside a larger agent. Its low latency makes it suitable for interactive features where a user is waiting, and its price makes it viable to run across large datasets.

It is not the model for your hardest problems. Competition-grade mathematics, deep multi-file refactoring and long autonomous agent runs are where full GPT-5’s extended reasoning earns its cost. The right pattern is to treat Mini as the workhorse and full GPT-5 as the specialist you call in when Mini’s output is not good enough — measured on your own evaluations rather than assumed.

Because Mini is part of the GPT-5 system it inherits the same safe-completions approach documented in the GPT-5 system card: helpful answers within safety boundaries rather than blunt refusals, with lower rates of hallucination, sycophancy and deception than the previous o-series and GPT-4o models. That makes it a safer default than o4-mini for user-facing features.

The usual caveats apply more strongly at the smaller size. Mini has less headroom for deep reasoning, so on ambiguous or knowledge-heavy queries it is more likely to need good retrieval and clear prompting to perform well. Evaluate it on representative tasks before relying on it, and build an escalation path to full GPT-5 for the cases it cannot handle.