GPT-5 Nano Guide: Reasoning at Five Cents Per Million Tokens

The interesting thing about GPT-5 Nano is not that it is cheap. Plenty of models are cheap. It is that it is a reasoning model at that price, which changes how you should budget for it. The input rate looks like a rounding error, but the bill is decided almost entirely by output, and output includes thinking you never see.

Nano launched with the rest of the GPT-5 family in August 2025. OpenAI calls it the fastest, most cost-efficient version of GPT-5 and names two jobs for it: summarisation and classification. That is a narrow brief, and the model is at its best when you respect it.

What Nano is, by the numbers

From OpenAI's model page:

• Price — $0.05 per million input tokens, $0.005 cached, $0.40 per million output
• Window — 400,000 tokens in total, with input capped at 272,000 and output at 128,000
• Modalities — text and image in, text out
• Knowledge cutoff — 31 May 2024
• Endpoints — Chat Completions, Responses and Batch, with streaming, function calling and structured outputs

The single snapshot is gpt-5-nano-2025-08-07. OpenAI's page now recommends GPT-5.6 Luna for most new speed- and cost-sensitive work, which tells you where the vendor's attention has moved, but Nano itself is still listed and still served.

A worked cost example

Take a support-ticket triage job: one million tickets a month, each prompt about 500 tokens including the instructions and the ticket, and a one-line label as the visible answer of around 20 tokens.

Now the part that matters. Suppose the model also spends 200 reasoning tokens per call. That figure is an assumption for illustration; the real number depends on your prompt and effort setting, and you can read it from usage.output_tokens_details.reasoning_tokens in the Responses API.

• Input: 500M tokens at $0.05 per million = $25
• Output: 220M tokens at $0.40 per million = $88
• Total: about $113 a month

Output is roughly three-quarters of the bill even though the visible answer is a single label. Cut reasoning to near zero and the job costs closer to $33. Let it double and you are past $190. The input price you saw on the pricing page is the least important number in the calculation.

The same job on GPT-5 Mini, at five times the rate on both sides, comes to roughly $565 with the same token counts. That gap is why Nano exists.

Keep the reasoning tax small

The GPT-5 family accepts minimal, low, medium and high reasoning effort, and the default is medium. For a classifier, medium is usually more thought than the task needs. Start at minimal, measure accuracy on a labelled sample, and only raise effort if the numbers say so.

Three other levers matter more on Nano than on bigger models:

1. Constrain the output. Use structured outputs with an enum of allowed labels. It removes a class of parsing failures and stops the model writing explanations you did not ask for.
2. Set verbosity low. The verbosity parameter controls the length of the final answer, independently of reasoning.
3. Cache the instructions. If every request shares a long system prompt, cached input at $0.005 per million makes it nearly free. Put the fixed part first and the variable ticket last.

If the taxonomy itself is fuzzy, no setting will save you. LLM classification makes the case for fixing the label set before blaming the model.

Where Nano breaks

Small models fail in predictable ways, and it is cheaper to know them in advance than to discover them in production.

• Multi-step instructions. A prompt with six conditional rules will see some of them dropped. Split it into separate calls, or move up a tier.
• Tool arguments that must be exact. Nano supports function calling, but a wrong ID or a malformed date in an argument costs more to catch than the tokens saved.
• Long-document reasoning. The 272K input cap lets you send a lot. It does not mean Nano reasons well across all of it. Summarising a long transcript works; answering a question that needs facts from three distant sections is where it slips.
• Anything after mid-2024. The knowledge cutoff is old. If the task leans on recent facts, they must be in the prompt.

Nano or Mini: a decision rule

Use Nano when the answer space is small and checkable: a label, a yes or no, a short summary, a routing decision. Use Mini when the model has to plan, chain a few steps, or produce arguments that downstream code trusts without checking.

When it is close, run both. Route everything to Nano first and escalate to Mini when a confidence check or validator fails. If fewer than one call in five escalates, that setup usually beats running Mini on everything. Model routing and fallbacks covers the pattern in detail.

Pinning a snapshot and running backfills

Pin gpt-5-nano-2025-08-07 rather than the bare alias for anything you have evaluated, so a future alias change cannot shift your accuracy numbers silently.

OpenRouter also lists openai/gpt-5-nano:batch, the same model on OpenAI's asynchronous batch endpoint at $0.025 input and $0.20 output per million tokens at the time of writing. A classification backfill rarely needs answers in seconds, so for bulk jobs the batch route halves an already small bill. For interactive triage, stay on the standard endpoint and accept the full rate.