GPT-5 Mini Guide: The 2025 Middle Tier, Priced for 2026

GPT-5 Mini shipped alongside GPT-5 in August 2025, and a year on it is easy to file it under "old". OpenAI's own model page now points new low-latency, high-volume work at GPT-5.6 Terra instead. But a lot of production traffic still runs on gpt-5-mini, and the question worth answering is not whether it is the newest model. It is whether it is still the right one for the job you have.

The short version: Mini is a real reasoning model at a price that makes reasoning affordable for bulk work, with a set of limits that are easy to misread. This guide covers what those limits are and how to decide between Mini, Nano and full GPT-5.

Where it sits in the GPT-5 family

OpenAI describes the GPT-5 family as three models: gpt-5, gpt-5-mini and gpt-5-nano. Mini is the middle one, described on the model page as a faster, more cost-efficient version of GPT-5 suited to well-defined tasks and precise prompts.

That phrasing is worth taking literally. Mini is not pitched at open-ended problems where the model has to work out what you meant. It is pitched at tasks where you already know what good output looks like and can say so in the prompt: extraction against a schema, rewriting to a style guide, answering from supplied documents, routine code edits with a clear target.

OpenAI's list prices per million tokens make the spacing clear:

• GPT-5 — $1.25 input, $0.125 cached input, $10 output
• GPT-5 Mini — $0.25 input, $0.025 cached input, $2 output
• GPT-5 Nano — $0.05 input, $0.005 cached input, $0.40 output

Each step down is a five-fold cut on both input and output. That ratio is the whole design: three models on one price curve, so you can move a workload up or down without changing anything else about your integration.

The 400K window is really 272K in

Catalogue listings, including OpenRouter's, show a 400,000-token context window and 128,000 tokens of maximum output. Both figures are correct, and both are easy to misuse.

OpenAI's model page gives the third number: a maximum input of 272,000 tokens. The 400K is the total budget for input plus output, and output includes hidden reasoning tokens. If you plan a job around stuffing 390K tokens of documents into one request, it will be refused before it starts.

The practical ceiling for a prompt is therefore 272K, and you should leave room inside the output budget for reasoning as well as the visible answer. There is a separate problem too: quality drops well before a window is full, and a smaller model is no exception. Test retrieval accuracy at the prompt sizes you actually plan to send.

Reasoning effort is the real price dial

Mini is a reasoning model. Before it writes the answer you see, it generates reasoning tokens you do not see, and those are billed at the output rate. On a model where output costs eight times input, that is where the money goes.

The GPT-5 family accepts four effort levels through reasoning.effort in the Responses API or reasoning_effort in Chat Completions: minimal, low, medium and high. The default is medium. OpenAI's guidance describes minimal as the fastest option that still keeps the benefits of reasoning, and suggests it for latency-sensitive work and for teams moving off GPT-4.1.

Two things follow. First, if you never set effort, you are paying for medium on every call, which is often more thinking than an extraction task needs. Second, at minimal the model plans less on its own, so OpenAI recommends prompting explicitly for planning and persistence. Effort and prompt are coupled; change one and re-test the other.

There is also a verbosity parameter, which controls the length of the final answer rather than the length of the thinking. Setting it low for terse structured responses is a cheap win that has nothing to do with reasoning depth, and it is worth doing before you touch effort at all.

When to pick Mini over Nano or full GPT-5

A decision rule that holds up in practice:

• Pick GPT-5 Nano when the task is classification, routing, short summaries or anything where the answer space is small. OpenAI positions Nano for exactly that, and it costs a fifth as much.
• Pick Mini when the task needs a few steps of reasoning, tool calls with arguments that have to be right, or reading a long document and answering from it. This is where Nano starts making errors you have to catch downstream.
• Pick full GPT-5 when the prompt cannot fully specify the task, when the work is multi-step coding, or when a wrong answer is expensive enough to dwarf a five-fold token premium.

The common mistake is choosing on per-token price alone. If Nano fails a fifth of the time and each failure triggers a retry or a human review, Mini can be cheaper per completed task. Measure that before committing a fleet to either.

Snapshots and the batch endpoint

OpenAI lists a single snapshot, gpt-5-mini-2025-08-07, as the default behind the gpt-5-mini alias. Pinning the dated name matters for anything you have evaluated: an alias can move, a snapshot does not.

OpenRouter also lists openai/gpt-5-mini:batch. It is the same model on the asynchronous batch endpoint, listed at $0.125 input and $1 output per million tokens at the time of writing, half the interactive rate. You give up immediate responses in exchange; results come back within the batch window. For nightly enrichment, backfills or evaluation runs, that trade is almost always worth taking. Batch API savings math covers when the discount genuinely beats a cheaper interactive model.

Should you start a new project on it

Probably not by default. OpenAI's own page steers new high-volume work to a newer model, and the knowledge cutoff of 31 May 2024 is old enough to matter for anything involving recent libraries or events.

Mini still makes sense in three cases: you already run it and have evaluations that pass, your workload is grounded entirely in supplied context so the cutoff does not matter, or you need a stable, well-understood model that will not change under you. If none of those apply, run the same twenty tasks through Mini and its newer replacement and compare cost per completed task. How to benchmark LLMs on your own work has the setup.