GPT-5.4 Nano Guide: The Cheapest 5.4 Model and Its Hard Limit

The cheapest model in a family is usually judged by one question: how much worse is it? For GPT-5.4 nano that question has two very different answers depending on the task, and getting the right one is most of the work of deploying it well.

OpenAI released nano alongside GPT-5.4 mini on 17 March 2026 and describes it as its cheapest GPT-5.4-class model for simple, high-volume tasks. It is available through the API only, not in ChatGPT. The current snapshot is gpt-5.4-nano-2026-03-17.

Two numbers that define nano

OpenAI's launch figures put nano at 52.4% on SWE-Bench Pro, against 54.4% for GPT-5.4 mini and 57.7% for the full GPT-5.4. On that benchmark it is two points behind mini and well ahead of the previous GPT-5 mini at 45.7%.

On OSWorld-Verified, which measures operating desktop software through screenshots, the picture changes completely. Nano scores 39.0%; mini scores 72.1% and GPT-5.4 scores 75.0%. That is not a small step down. It is a different class of capability.

OpenAI's API page reflects the same split. Mini and GPT-5.4 both list computer use and tool search among their Responses API tools. Nano lists function calling, web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills and MCP, but not computer use or tool search. The lesson: nano can handle bounded, text-shaped problems surprisingly well, and should not be put in charge of multi-step interaction with an interface.

The jobs it is built for

OpenAI positions nano for classification, data extraction, ranking and simpler supporting subagent work. Those tasks share a shape: a fixed instruction, a short variable input, and a small, constrained output that code can validate.

• Routing and triage. Tagging tickets, deciding which queue or which model a request should go to, flagging likely spam.
• Extraction. Pulling fields from invoices, emails or logs into a schema, with structured outputs enforcing the shape.
• Ranking and filtering. Scoring search results or retrieved chunks for relevance before a larger model reads them.
• Cheap checks inside a pipeline. Does this diff touch a migration? Does this answer cite a source? Yes or no, at high volume.

The common thread is that the taxonomy and the schema do most of the work. LLM classification covers why fixing the label set matters more than the model behind it.

What it costs at volume

OpenAI prices nano at $0.20 per million input tokens, $0.02 for cached input and $1.25 per million output, and OpenRouter lists the same rates at the time of writing. Per-token prices this small are hard to reason about, so put them in a real job.

Take one million classifications, each with a 2,000-token prompt (instructions plus a few examples), of which 1,800 tokens are the same every time, 200 tokens of variable input, and a 20-token answer.

• Nano, no caching: 2 billion input tokens at $0.20 is $400, plus $25 of output. About $425.
• Nano, shared prefix cached: 1.8 billion cached tokens at $0.02 is $36, 200 million fresh input tokens is $40, plus $25 of output. About $101.
• Mini, shared prefix cached: $135 cached, $150 fresh input and $90 output. About $375.

Notice that mini with caching is cheaper than nano without it. Prompt structure moved the bill more than the choice of model did. Put the fixed instructions and examples first, keep them byte-identical across requests, and put the variable input last. Providers only cache prefixes above a minimum length, so check that your shared section qualifies. Prompt caching savings walks through the break-even arithmetic.

The batch endpoint halves it again

Most nano workloads do not need an answer in under a second. Backfilling labels on a year of tickets or re-scoring a document corpus can wait an hour. OpenAI supports its Batch endpoint for nano, and OpenRouter lists openai/gpt-5.4-nano:batch, the same model on an asynchronous endpoint at $0.10 input and $0.625 output per million tokens at the time of writing, half the standard rate.

The trade is latency for price, plus a little operational work: you submit a file of requests and collect the results later. Batch API savings covers when that trade pays off and when the plumbing costs more than it saves.

Limits worth knowing before you switch

The context window is listed at 400,000 tokens, but input is capped at 272,000 with up to 128,000 tokens of output. Knowledge cutoff is 31 August 2025. Reasoning effort defaults to none and can be raised through low, medium, high and xhigh.

That last setting deserves care. Raising effort on nano spends hidden output tokens, and output is the expensive side of its pricing. For a classification job, a higher effort level can easily multiply the output bill several times over while the answer stays one word long. Test whether it changes accuracy on your labelled set before you turn it on.

When to step up to mini

Move a task from nano to GPT-5.4 mini when any of these is true:

• The task involves computer use or navigating an interface over several steps.
• The output is open-ended rather than a label, a score or a schema.
• Your evaluation shows nano failing on more than a small share of cases, and failures are not caught automatically.
• You expose a large tool catalogue and want tool search to keep definitions out of the prompt.

Otherwise, nano is the right default for the bottom of the stack. Build a labelled set of a few hundred real examples, measure accuracy at effort none, cache the shared prefix, and send anything that can wait to the batch endpoint. When a cheap model is enough covers how to set the accuracy bar before you start.