Gemini 3.1 Pro is Google's top Gemini model for developers: it reads up to a million tokens of text, images, audio, video and PDFs and reasons over them. It is still a preview model, with no free API tier, at $2 per million input tokens. Pick it for hard analysis of long material; use Flash models for routine volume.
Other good fits
Harder reasoning inside an app or automation, where a cheaper Flash model gives shallow or wrong answers.
+ Pros
Read pros
Gemini 3.1 Pro is the most capable model in Google's Gemini API, made by Google DeepMind and released in February 2026. Its model ID is gemini-3.1-pro-preview: seven months on, it is still labelled preview. It replaced Gemini 3 Pro Preview, which Google shut down on 9 March 2026.
Its strength is breadth of input. One request can hold up to about a million tokens, roughly 1,500 pages of text, and mix PDFs, images, audio and video. It can also call your functions, return structured JSON, and ground answers in Google Search.
Pick it when:
The answer depends on reading everything: comparing ten supplier contracts, finding contradictions across a thesis and its data appendix, summarising a full day of meeting recordings.
A Flash model's answers are too shallow for the task and you've checked that on real examples.
You want Gemini's strongest model inside your own app, script or automation, not only in the chat window.
− Cons
Read cons
Preview status. Behaviour and availability can change. For a system you'll depend on for a year, build in a way to switch models.
Long context isn't perfect recall. With hundreds of pages, details from the middle get missed. Ask for quotes and page numbers and spot-check them.
No free API tier for this model, unlike the Flash models.
Your data goes to Google's cloud. Check with your data protection officer before sending personal or confidential material.
Pick something else when:
The job is routine extraction or classification: Gemini Flash-Lite costs about an eighth as much.
Your organisation already buys another vendor: the Claude API or GPT-5 are like-for-like choices for long-document reasoning.
Documents may not leave your building: run an open model such as Gemma 4 locally.
Avoid if
The task is simple and repetitive. Gemini Flash-Lite does it at about an eighth of the price.
You only want to chat with it. The Gemini app gives you 3.1 Pro without any code.
Context
The catalogue name is Gemini 3.1 Pro. Confirm its exact current API identifier, context limit and supported modalities in Google's live model documentation before implementation.
Requirements
Use a supported Google API project, authenticated client and the selected endpoint.
Licence & cost
Read licence & cost
Prices checked 23 September 2026 (Gemini API, USD per million tokens, paid tier only):
Prompt size
Input
Output (incl. thinking)
Up to 200,000 tokens
$2.00
$12.00
Over 200,000 tokens
$4.00
$18.00
The Batch API roughly halves those prices for jobs that can wait. Thinking tokens count as output, so a hard question can cost more than its short answer suggests. Paid-tier content is not used to improve Google's products.
A 100-page PDF comparison typically costs cents, not euros. A daily job over large archives adds up quickly; set a budget alert on the billing account and log the token count per run.
In the app, 3.1 Pro comes without any API setup: limited on the free plan, full on Google AI Pro at €21.99 a month.
Operating effort: keep the API key out of shared files, write the model ID in one place, and re-check a few outputs whenever Google moves the model. Preview models can change or be retired at shorter notice than stable ones; Google says at least two weeks.
First task
Documentation-based suggestion — not run here.
Read first task
Compare three long documents and have the model show its sources. Allow 45 minutes; you need a Google account with billing enabled in Google AI Studio (opens in a new tab), because 3.1 Pro has no free API tier.
Pick three public documents that cover the same topic, for example three funders' grant guidelines or three versions of a safety standard.
In AI Studio, choose Gemini 3.1 Pro and upload all three PDFs.
Ask: "Build a table of every requirement that differs between these documents. For each row give the document, page number and exact quote. Mark anything you're unsure of."
Check five random rows against the PDFs.
Success check: all five checked rows have the right quote on the stated page, and the table surfaces at least one difference you hadn't noticed. Note the token count AI Studio shows; it tells you what the same job would cost through the API. T04-L03 covers answers that cite their sources.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Checked against Google's Gemini API model list, the 3.1 Pro model page, pricing page, deprecation schedule, API terms and the Gemini app subscriptions page on 23 September 2026. We did not send API requests for this review.
Gemini API pricing · Sep 23, 2026 — 3.1 Pro Preview: $2.00 input / $12.00 output per 1M tokens up to 200k-token prompts, $4.00 / $18.00 above; no free tier.
Gemini 3.1 Pro model page · Sep 23, 2026 — Model ID gemini-3.1-pro-preview; 1,048,576 input and 65,536 output tokens; text, image, video, audio and PDF input; latest update February 2026.
Gemini deprecations · Sep 23, 2026 — gemini-3.1-pro-preview released 19 February 2026, no shutdown date announced; gemini-3-pro-preview shut down 9 March 2026.
Gemini models · Sep 23, 2026 — 3.1 Pro listed under Preview; 3.8 Flash is the newest stable model.
Gemini subscriptions · Sep 23, 2026 — Free app plan has limited 3.1 Pro; Google AI Pro (€21.99/month) gives full access.
Further review notes & alternatives
Alternatives and trade-offs
Claude API wins for careful long-form writing and editing, and for teams already using Claude.
GPT-5 wins when you're already on OpenAI and want one vendor for chat, API and agents.
Gemini Flash-Lite wins when you have thousands of simple items and cost per item is what matters.
Analyze large documents, media files, and whole code repositories.
Use multimodal prompts with a context window suited to long inputs.
Choose Deep Think when a difficult problem needs extra reasoning.
