Gemini 3.1 Pro Aces Benchmarks, I Suppose
The new model scores 77% on ARC-AGI-2, a 2x improvement over Gemini 3 Pro, and leads multiple independent leaderboards.
Google DeepMind has launched Gemini 3.1 Pro, positioning it as a significant step forward in core reasoning capabilities. Announced by CEO Sundar Pichai, the model scores 77.1% on the ARC-AGI-2 benchmark, representing a more than 2x improvement over Gemini 3 Pro. Google is shipping the model across its consumer and developer ecosystems immediately, aiming to bring enhanced intelligence to everyday applications. The release focuses on the model's strength in handling super-complex tasks like data synthesis, creative projects, and visualizing difficult concepts, backed by demonstrations in urban planning simulation and heat transfer analysis.
Independent benchmark aggregators confirm Gemini 3.1 Pro's strong performance. Artificial Analysis shows it leading by three full points, while LiveBench places it at 79.93, 3.6 points ahead of Claude Opus 4.6. It also dominates the Voxelbench (1725 vs. GPT-5.2's 1531) and Mercor's APEX-Agents leaderboards, completing five previously unsolved tasks. The model card highlights its suitability for agentic performance, advanced coding, and long-context understanding. However, the announcement provides only a brief summary of its frontier safety test results, noting it reaches the same 'alert' thresholds as Gemini 3 Pro without new concerns, a transparency level some critics find lacking for a frontier model candidate.
- Scores 77.1% on ARC-AGI-2, a 2x+ improvement over Gemini 3 Pro in core reasoning.
- Leads independent benchmarks like LiveBench (79.93 score) and Voxelbench, beating Claude Opus 4.6 and GPT-5.2.
- Optimized for agentic tasks, advanced coding, and long-context understanding, now available across Google's product suite.
Why It Matters
Establishes a new performance benchmark for reasoning AI, directly impacting developers building complex agents and enterprise applications.