Gemini 3.1 Pro review: accurate, but watch the thinking tokens
Gemini 3.1 Pro (gemini-3.1-pro-preview) made no clear mistakes on any task we gave it, and on the one task built to trip models up, it noticed the trap. The catch is cost shape: it is a reasoning model, and even a short answer comes with well over 1,000 tokens of thinking that you pay for as output.
What it is
Gemini 3.1 Pro is Google's top reasoning model, released as a preview in February 2026.
- Context: about 1M tokens, with up to 65,536 output tokens
- Input: text, images, audio, video and PDF
- Output: text
- Thinks internally before every answer
Five tasks, measured
We sent the same five Japanese-language tasks to several models. These tasks are useful precisely because they punish sloppy models: politeness levels, seasonal words in poetry and arithmetic consistency all have right and wrong answers. Gemini 3.1 Pro's results:
| Task | Result | Output tokens (of which thinking) |
|---|---|---|
| Apology email for a delayed delivery | Natural tone; also suggested listing the old and new dates side by side | 1,719 (1,489) |
| Three-point summary | Accurate, with readable headings | 915 (742) |
| Fix two honorific errors | Correct, with the right reason for each (humble vs respectful forms) | 1,271 (1,021) |
| Arithmetic word problem | Spotted that the numbers in the problem contradict each other | 2,393 (2,087) |
| Haiku with a seasonal word | Correct 5-7-5 meter and a correct summer seasonal word | 2,502 (2,439) |
The word problem was deliberately unsolvable: worked through, it gives 7.2 apples. Gemini 3.1 Pro solved it with the classic two-unknowns method, then said the problem's numbers must be wrong, and proposed a fix: change the stated total from 1,215 to 1,125 and the answer becomes a clean 6. Some of the other models we tested gave a confident, plausible and wrong answer to the same problem.
The haiku task is one where models often get the season of the seasonal word wrong. Gemini wrote:
短夜や 祈るビルドに 射す朝日
Roughly: "Short summer night / on the build we pray for / the morning sun falls." The meter, the season (summer) and its own explanation of the poem all held up.
Reading a receipt image
We gave it a photo of a Japanese restaurant receipt and asked for JSON with the store name, date and time, line items, which items fall under Japan's reduced 8% tax rate, and the store's invoice registration number. Every field came back correct, including the items marked with ※ as reduced-rate. The image cost about 1,100 input tokens.
Cost: look at thinking tokens before the rate
In the table above, a roughly 250-character email produced about 1,700 output tokens, and about 90% of them were thinking. Thinking tokens are billed as output, so the shorter the answer, the larger the share of the bill that thinking takes.
- What one call costs depends on how much the model thought, more than on how long the answer is
- Prompts over 200K tokens move to a higher input and output rate, following Google's own tiering
- For high-volume routine work, pair it with a lighter model
Gemini 3.1 Pro is billed per million input and output tokens; current rates are on the model page and on /pricing.
Where it fits
Good fit:
- Writing and proofreading where precise wording matters
- Checks where the value is noticing that something is wrong, such as numbers that do not add up
- Analysis over long documents, images and PDFs together
- A small number of heavy decisions, such as planning steps for an agent
Poor fit:
- Chat that must answer within a few seconds (our calls took 10 to 20-odd seconds each)
- Large volumes of short classification or extraction, where thinking overhead dominates
FAQ
Is the preview safe to use in production?
Preview terms can change. For anything important, keep your code able to swap the model value without touching anything else, so you can move to another model on the same key.
Can I turn the thinking off?
No, but you can set its depth with reasoning.effort. See reasoning control.
Does Gemini 3.1 Pro take images?
Yes. Send images alongside text; it read a Japanese receipt accurately in our test.
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": "gemini-3.1-pro-preview",
"reasoning": {"effort": "low"},
"messages": [{"role": "user", "content": "A store sells apples and oranges. Check whether the numbers in this problem are consistent before solving it: 10 fruits cost 1,215 in total, apples cost 150 each and oranges cost 100 each. How many apples?"}]
}'
The same key reaches every other model in the catalog, so you can run the same prompt against Claude or a lighter Gemini by changing model.