GPT-4.1 nano review: 1M context, cheap image extraction, shaky reasoning
GPT-4.1 nano (gpt-4.1-nano) is for running large volumes of light work cheaply and fast. In our tests its emails and summaries were usable and its image extraction was accurate. Tasks that need it to think, like explaining a correction or solving a word problem, came out rough. Use it for routine processing that involves no judgment.
Specs
GPT-4.1 nano is the smallest and cheapest model in OpenAI's GPT-4.1 series, released in April 2025.
- Context: about 1M tokens, up to 32,768 tokens of output
- Input: text, images and files
- Output: text
- Tool calls and structured output (JSON Schema)
It is billed per million input and output tokens; current prices are on the model page.
Five writing and reasoning tasks
We sent the same five tasks, all written in Japanese, to several models. GPT-4.1 nano answered each in about 1 to 5 seconds, among the fastest in the group.
| Task | Result |
|---|---|
| Apology email for a late delivery | Natural and usable, though it used a formal letter closing without the matching opening |
| Three-point summary | Accurate and the most concise of the group (97 output tokens) |
| Fix two honorific errors | The corrected sentence was right, but its explanation contradicted itself along the way |
| Math word problem | Noticed there was no solution, but switched to English partway through |
| Haiku with a seasonal word | Not 5-7-5, and mixed an autumn and a spring seasonal word |
The word problem was built so that no integer solution exists. GPT-4.1 nano deserves credit for spotting that, but it answered a Japanese question in English. If you need output in a specific language, say so in the system prompt.
Image extraction: cheap and accurate
We gave it a photo of a Japanese store receipt and asked for the store name, date and time, line items, which items fell under the reduced tax rate, and the invoice registration number, as JSON. It read every field correctly, including the items marked with an asterisk as reduced-rate.
The input token count is the real story. The same image cost GPT-4.1 nano 1,220 input tokens; OpenAI's GPT-4o mini took 25,568, about 20 times as many. These counts were measured in May 2026; GPT-4o mini has since moved to a different host on FastMetal, so re-measure before relying on the exact ratio. For high-volume image extraction, nano is a strong option on cost.
On a simple test asking it to describe where solid-colored shapes sat in an image, it sometimes got the positions wrong. It reads text in images well; fine-grained understanding of diagrams is not its strength.
Good fit, poor fit
Good fit
- Extraction from text-heavy images such as invoices and receipts
- Summarizing or searching long documents, up to about 1M tokens
- High-volume classification, tagging and templated text
Poor fit
- Explaining the reasoning behind a correction or edit
- Arithmetic, or spotting contradictions in the conditions of a problem
- Detailed reading of charts and diagrams
FAQ
Q. How does GPT-4.1 nano compare with GPT-4o mini?
GPT-4o mini's writing is slightly steadier. GPT-4.1 nano used about a twentieth of the input tokens for the same image and has a longer context, so it wins on high-volume and long-document work.
Q. Why does GPT-4.1 nano sometimes answer in English?
We saw it switch to English mid-answer on a math problem. Naming the response language in the system prompt prevents it.
Q. Does GPT-4.1 nano support structured output?
Yes. With a JSON Schema in response_format, it returned JSON that matched the schema.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-4.1-nano",
"messages": [
{"role": "system", "content": "Always answer in English."},
{"role": "user", "content": "Summarize in three bullet points: Small businesses are adopting generative AI, but unpredictable costs remain a barrier. Bills in a foreign currency move with exchange rates and can come in higher than planned."}
]
}'
The same API key and balance work across the catalog, so you can change model and compare. Prices are on the pricing page and every model is in the model catalog.