Google's Gemini 3.1 Pro Achieves 77.1% on ARC-AGI-2, Doubling Previous Reasoning Capabilities
What is Gemini 3.1 Pro?
On February 19, 2026, Google unveiled Gemini 3.1 Pro, the latest addition to its Gemini 3 series. As the successor to Gemini 3 Pro, it brings substantial enhancements to reasoning performance, multimodal support, and agent functionalities.
Key Specifications
| Item | Details |
|---|---|
| Context Window | 1M tokens |
| Max Output | 64K tokens |
| Supported Inputs | Text, Code, Images, Audio, Video, PDF |
| Knowledge Cutoff | January 2025 |
| Availability | API (Preview), Gemini App, NotebookLM |
Main Benchmark Results
Reasoning
| Benchmark | Gemini 3.1 Pro | Gemini 3 Pro | Claude Opus 4.6 | GPT-5.2 |
|---|---|---|---|---|
| ARC-AGI-2 | 77.1% | 31.1% | 68.8% | 52.9% |
| GPQA Diamond | 94.3% | 91.9% | 91.3% | 92.4% |
| HLE (Academic Reasoning) | 44.4% | 37.5% | 40.0% | 34.5% |
On ARC-AGI-2, it surged from 31.1% to 77.1%, marking a 2.5x improvement in the ability to solve entirely novel logical patterns.
Coding and Agents
| Benchmark | Gemini 3.1 Pro | Claude Opus 4.6 | GPT-5.3-Codex |
|---|---|---|---|
| SWE-Bench Verified | 80.6% | 80.8% | — |
| SWE-Bench Pro | 54.2% | — | 56.8% |
| Terminal-Bench 2.0 | 68.5% | 65.4% | 64.7% |
| SciCode | 59% | 52% | — |
| LiveCodeBench Pro | 2887 Elo | — | — |
It nearly matches Claude Opus 4.6 (80.8%) on SWE-Bench Verified and outperforms it on Terminal-Bench 2.0 with 68.5% versus 65.4%.
Multimodal and Multilingual
| Benchmark | Gemini 3.1 Pro | Gemini 3 Pro |
|---|---|---|
| MMMU-Pro | 80.5% | 81.0% |
| MMMLU | 92.6% | 91.8% |
| MRCR v2 (128K) | 84.9% | 77.0% |
Agent and Tool Usage
| Benchmark | Gemini 3.1 Pro | Claude Opus 4.6 | GPT-5.2 |
|---|---|---|---|
| MCP Atlas | 69.2% | 59.5% | 60.6% |
| BrowseComp | 85.9% | 84.0% | 65.8% |
| τ2-bench (Retail) | 90.8% | 91.9% | 82.0% |
It significantly outperforms Claude Opus 4.6 (59.5%) on MCP Atlas with a 69.2% score.
Strengths of Gemini 3.1 Pro
1. Dramatic Improvement in Abstract Reasoning
The 77.1% on ARC-AGI-2 is a leap from the previous generation's 31.1%. This benchmark evaluates the ability to tackle "completely new logical patterns," serving as a key indicator of a model's generalization capability.
2. Enhanced Agent Workflows
Scores like MCP Atlas at 69.2% and BrowseComp at 85.9% highlight superiority in multi-step tool usage tasks. Google emphasizes improvements in agent capabilities for domains such as "finance and spreadsheet applications."
3. Expanded Thinking Levels
A new MEDIUM thinking level has been introduced, allowing for finer adjustments in the trade-off between cost, performance, and speed. Combined with the existing HIGH, it simplifies optimization for various use cases.
4. Custom Tool Endpoints
A dedicated endpoint, gemini-3.1-pro-preview-customtools, is now available, optimized for agent workflows that prioritize custom tools like view_file or search_code.
Pricing and Availability
Gemini 3.1 Pro is accessible on the following platforms:
- For Developers: Gemini API (Google AI Studio), Gemini CLI, Google Antigravity, Android Studio
- For Enterprises: Vertex AI, Gemini Enterprise
- For Consumers: Gemini App, NotebookLM (Pro/Ultra plans)
Subscribers to Google AI Pro ($20/month) and Ultra ($100/month) plans can access Gemini 3.1 Pro in the Gemini app with enhanced usage limits.
Positioning Compared to Other Models
In the early 2026 AI model landscape, Gemini 3.1 Pro stands out as follows:
| Model | Strengths | GPQA Diamond | SWE-Bench Verified |
|---|---|---|---|
| Gemini 3.1 Pro | Abstract Reasoning, Multimodal | 94.3% | 80.6% |
| Claude Opus 4.6 | Coding, Long-context Processing | 91.3% | 80.8% |
| Qwen3.7-Max | Agent, Long-term Autonomy | 92.4% | 80.4% |
| GPT-5.2 | Balanced | 92.4% | 80.0% |
Its GPQA Diamond score of 94.3% is the highest among published results.
Summary
Gemini 3.1 Pro represents a major upgrade in reasoning capabilities from Gemini 3 Pro. The standout 77.1% on ARC-AGI-2 demonstrates exceptional generalization for abstract logical problems, surpassing other models significantly.
Key highlights:
- ARC-AGI-2: 77.1% (2.5x improvement over the previous generation)
- GPQA Diamond: 94.3% (highest recorded score)
- SWE-Bench Verified: 80.6% (on par with Claude Opus 4.6)
- MCP Atlas: 69.2% (leads in agent capabilities)
- 1M tokens context window
- Addition of MEDIUM thinking level for streamlined cost optimization
Loading...