GPT-4o mini review: steady writing, heavy image tokens
GPT-4o mini (gpt-4o-mini) is a steady lightweight model: emails and summaries come out clean and you can leave them to it. Two things to watch. It can miss a contradiction in the numbers and give a plausible wrong answer anyway. And image input is expensive in tokens: one photo used over 25,000 input tokens in our test, which changes the economics of any image workload.
Specs
OpenAI released GPT-4o mini in July 2024.
- Context: 128,000 tokens, up to 16,384 tokens of output
- Input: text and images
- 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-4o mini answered each in about 2 to 7 seconds, which feels quick next to models that reason before answering.
| 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 concise |
| Fix two honorific errors | The corrected sentence was right; the explanation used an unnatural honorific form |
| Math word problem | Wrong. It noticed midway that the answer came to 7.2 apples, then concluded "9 apples" anyway |
| Haiku with a seasonal word | 締め切りの 夢に追われて 秋深し ("chased through dreams by a deadline; deep autumn"). Correct 5-7-5, sensible seasonal word |
The word problem was built so that no integer solution exists. GPT-4o mini spotted the contradiction, then gave "9" with nothing to support it, and 9 apples does not match the total in the problem. Do not use it alone for anything where the numbers have to be checked; add a verification step.
Image extraction: accurate, and token-hungry
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. Every field came back correct, including the reduced-rate items.
The token count is the catch. That one image cost GPT-4o mini 25,568 input tokens. OpenAI's GPT-4.1 nano read the same image for 1,220. OpenAI meters GPT-4o mini's image input at a high token count by design, so a model that is cheap per text token can still be expensive per image.
These counts were measured in April 2026; GPT-4o mini has since moved to a different host on FastMetal. If you process images at volume, send a sample image to each candidate model and compare the usage.prompt_tokens before you choose. Our GPT-4.1 nano review has the other side of this test.
Good fit, poor fit
Good fit
- Emails, templated text and summaries
- Lightweight apps built on tool calls or JSON output
- Chat where response speed shapes the experience
Poor fit
- High-volume image pipelines (the token count adds up)
- Arithmetic or numeric consistency checks without a verification step
FAQ
Q. GPT-4o mini or GPT-4.1 nano?
GPT-4o mini if you want the steadier writer. GPT-4.1 nano if you handle many images or want to spend less: on the same image it used 1,220 input tokens against GPT-4o mini's 25,568.
Q. Is its writing natural?
Yes for emails and summaries. Formal conventions, such as matching a letter's opening and closing, are safer spelled out in the prompt.
Q. Does GPT-4o mini 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-4o-mini",
"messages": [{"role": "user", "content": "Write a 100-word apology email to a client for a delayed delivery."}]
}'
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.