Gemini 3.1 Pro
Gemini 3.1 Pro is Google's preview reasoning model for advanced coding, agentic workflows, multimodal understanding, and complex tasks. It supports up to 1M input tokens and 64K text output.
GoogleGemini 3.1 Pro release date and key capabilities
Gemini 3.1 Pro is most compelling when a task combines difficult reasoning with code, tools, or multimodal context. It is less about routine chat and more about complex work that benefits from a larger reasoning budget.
Advanced coding and agents
Use Gemini 3.1 Pro for repository work, terminal tasks, multi-step tool calls, and agent workflows that must keep making progress.
Complex reasoning
The model improves sharply over Gemini 3 Pro on abstract reasoning and long-horizon professional-task benchmarks.
Multimodal synthesis
Combine text, code, images, video, audio, and PDFs when a task depends on understanding more than plain text.
Official benchmark results for Gemini 3.1 Pro
Google DeepMind reports these scores for Gemini 3.1 Pro Thinking (High). Harnesses and settings matter, so treat them as model-selection signals rather than guarantees.
The official February 2026 model card compares Gemini 3.1 Pro with Gemini 3 Pro, Sonnet 4.6, Opus 4.6, GPT-5.2, and GPT-5.3-Codex. It does not publish a direct Gemini 3.1 Pro vs GPT-5.1 table.
A multimodal reasoning model for tools and long context
Gemini 3.1 Pro builds on Gemini 3 Pro and keeps the same broad multimodal and tool-use foundation while improving reasoning and agentic performance.
Google does not publish a parameter count for Gemini 3.1 Pro. The official model card describes it as the next iteration of Gemini 3 Pro, with native multimodal inputs, text output, a 1M-token context window, and support for tool-driven workflows.
Native multimodal input
Reason across text, code, images, video, audio, and PDF files in one prompt.
1M-token context
Bring large repositories, long documents, recordings, and task history into a single session.
Tool-ready output
Use function calling, structured output, search, and code execution in agent workflows.
Controlled reasoning
Thinking settings let developers balance deeper analysis against latency and cost.
Gemini 3.1 Pro context window and agent performance
The 1M context window is useful for large inputs, but context capacity is not the same as perfect retrieval. Google reports 84.9% on MRCR v2 at 128K average and 26.3% at the 1M pointwise test, so production teams should still test retrieval quality on their own data.
Gemini 3.1 raises the bar across a wide range of benchmarks
Compare Gemini 3.1 Pro with Gemini 3 Pro, Sonnet 4.6, Opus 4.6, GPT-5.2, and GPT-5.3-Codex across reasoning, coding, agents, tools, search, multimodal understanding, and long context.
The clearest gain is abstract reasoning
Gemini 3.1 Pro reaches 77.1% on ARC-AGI-2 versus 31.1% for Gemini 3 Pro, the largest generation-over-generation jump in the official table.
Gemini 3.1 Pro vs Opus 4.6 is workload-dependent
Gemini 3.1 Pro leads under the common Terminal-Bench harness and on LiveCodeBench Pro, but Opus 4.6 is slightly ahead on SWE-Bench Verified and GPT models lead SWE-Bench Pro.
A 1M window still needs retrieval testing
Gemini 3.1 Pro supports 1M input tokens, yet Google's pointwise MRCR score at 1M is 26.3%. Capacity alone does not guarantee reliable recall.
| Performance | Benchmark notes | Gemini 3.1 Pro Thinking (High) | Gemini 3 Pro Thinking (High) | Sonnet 4.6 Thinking (Max) | Opus 4.6 Thinking (Max) | GPT-5.2 Thinking (xhigh) | GPT-5.3-Codex Thinking (xhigh) |
|---|---|---|---|---|---|---|---|
| Humanity's Last Exam | Academic reasoning, full text + multimodal set, no tools | 44.4% | 37.5% | 33.2% | 40.0% | 34.5% | — |
| ARC-AGI-2 | Abstract reasoning puzzles, ARC Prize Verified | 77.1% | 31.1% | 58.3% | 68.8% | 52.9% | — |
| GPQA Diamond | Scientific knowledge, no tools | 94.3% | 91.9% | 89.9% | 91.3% | 92.4% | — |
| Terminal-Bench 2.0 | Agentic terminal coding, Terminus-2 harness | 68.5% | 56.9% | 59.1% | 65.4% | 54.0% | 64.7% |
| SWE-Bench Verified | Agentic coding, single attempt | 80.6% | 76.2% | 79.6% | 80.8% | 80.0% | — |
| SWE-Bench Pro (Public) | Diverse agentic coding tasks, single attempt | 54.2% | 43.3% | — | — | 55.6% | 56.8% |
| LiveCodeBench Pro | Competitive coding problems, Elo | 2887 | 2439 | — | — | 2393 | — |
| APEX-Agents | Long-horizon professional tasks | 33.5% | 18.4% | — | 29.8% | 23.0% | — |
| MCP Atlas | Multi-step workflows using MCP | 69.2% | 54.1% | 61.3% | 59.5% | 60.6% | — |
| BrowseComp | Agentic search with Search + Python + Browse | 85.9% | 59.2% | 74.7% | 84.0% | 65.8% | — |
| MMMU-Pro | Multimodal understanding and reasoning, no tools | 80.5% | 81.0% | 74.5% | 73.9% | 79.5% | — |
| MRCR v2 (8-needle) | Long-context performance, 128K average | 84.9% | 77.0% | 84.9% | 84.0% | 83.8% | — |
Scores and configurations follow Google DeepMind's Gemini 3.1 Pro Performance table. A dash means Google did not publish a score in this comparison. Vendor-reported benchmark settings can differ, so test your own workload before choosing a model.
Where Gemini 3.1 Pro fits
Large repository work
Review code, issues, logs, and specifications together before planning or implementing a change.
Research and synthesis
Use multimodal inputs, search, and code execution to turn complex sources into a structured result.
Agentic coding
Test multi-step terminal, MCP, and browser workflows where the model must plan, act, and revise.
Cross-model evaluation
Compare Gemini 3.1 Pro with Opus, Sonnet, and GPT models on your own prompts, tools, latency targets, and long-context retrieval needs.
Gemini 3.1 Pro FAQs
Gemini 3.1 Pro
Evaluate Gemini models in a hosted MyClaw agent
Test supported Gemini models with files, tools, browser-ready execution, and long-context agent workflows in one managed workspace.