Model cards

gemini-3.1-pro

Vendor google Tier - Price per million tokens, in / out $2 / $12 Released 2026-02-19 Context 1,000,000 Max output 64,000 Vendor page

Last updated 2026-10-26 How this is made

Meant to excel at

• Built for tasks where a simple answer is not enough; ARC-AGI-2 verified 77.1%, more than double Gemini 3 Pro.

  Vendor's own site blog.google

• Showcased for visual explanations, code-to-animated-SVG, live API dashboards, interactive 3D and creative coding.

  Vendor's own site blog.google

• 1M-token context with text, image, audio and video input; fast generation (about 139 tokens/s) with relatively concise output.

  Independent evaluator artificialanalysis.ai

• Handles long dense documents without truncation and finds details across large document sets that keyword search misses.

  Independent evaluator layer3labs.io

• One reviewer found it stayed on task per prompt and was better at Swift and native Apple APIs than Opus.

  Independent evaluator getneotiler.com

Watch out for

• Slow to start: about 19.8s time to first token versus a 3.9s median.

  Independent evaluator artificialanalysis.ai

• Mid-pack intelligence (index 30, rank 98 of 217) with above-median output price ($12 per 1M).

  Independent evaluator artificialanalysis.ai

• Can fabricate citations, DOIs and author lists; research output needs verification before use.

  Independent evaluator layer3labs.io

• A practitioner says it drifts off task in long-horizon agentic work, unlike Opus or GPT-5.5 class models.

  Practitioner report news.ycombinator.com

Benchmarks

Agentic coding percentile 0.43

Overall percentile 0.04

Reasoning percentile 0.67

Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.

Practitioner reports

• +

  Handles larger codebases better than Gemini 3.8 Flash: a question answered in about a minute by 3.1 Pro took 15 to 20 minutes with Flash.

  Compared with gemini-3.8-flash | 2026-10-14

  Practitioner report discuss.ai.google.dev
• +

  An Ultra user stayed on 3.1 Pro because 3.8 Flash failed after 20+ attempts at feature work in an existing codebase.

  Compared with gemini-3.8-flash | 2026-10-14

  Practitioner report discuss.ai.google.dev
• +

  Subscribers complain Google removed 3.1 Pro from Antigravity for Pro users, leaving only Flash, and say Flash cannot replace a Pro model.

  Compared with gemini-3.8-flash | 2026-10-14

  Practitioner report discuss.ai.google.dev
• +/-

  Judged better than Opus at Swift and Apple APIs and free of the context-bleed seen in Qwen 3.8 Max, but weaker at structured planning.

  Compared with qwen3.8-max | 2026-09-16

  Practitioner report getneotiler.com
• +/-

  Useful for biomedical literature screening and assay-parameter extraction, but it can invent citations, DOIs and journals.

  2026-09-09

  Practitioner report layer3labs.io
• -

  A July 2026 decision guide advises against using 3.1 Pro for general coding and recommends a Flash model as the coding default; keeps 3.1 Pro only as the long-context specialist (author did not test it first-hand).

  Task: feature | Compared with gemini-3.6-flash | 2026-07-22 older than 90 days

  Practitioner report aimadetools.com
• +/-

  A builder-oriented comparison rates 3.1 Pro behind GPT-5.5 and Opus 4.7 for code and agentic reliability, with its 1M context and multimodal strength as the main edge. NOTE: April 2026, outside 120-day window.

  Task: agentic-long-running | Compared with claude-opus-4.7 | 2026-04-26 older than 90 days

  Practitioner report mindstudio.ai
• -

  Says Gemini 3.1 Pro drifts off task on long-horizon agentic work, unlike Opus 4.6+, GPT-5.5 and GLM-5.2.

  Compared with glm-5.2 undated

  Practitioner report news.ycombinator.com