Gemini 3.1 Pro Doubles Reasoning Scores, Beats Claude and GPT on ARC‑AGI‑2

Benchmark Highlights

Google DeepMind released official numbers showing Gemini 3.1 Pro reaches 77.1% on ARC‑AGI‑2, a 148% increase over the previous 31.1% version. It also scores 100% on the AIME 2025 mathematics competition, 80.6% on SWE‑Bench Verified, 2887 Elo on LiveCodeBench Pro, and 94.3% on GPQA Diamond.

Comparison with Competitors

Against Claude Opus 4.6 (68.8% on ARC‑AGI‑2) and GPT‑5.2 (52.9%), Gemini 3.1 Pro leads in abstract reasoning. On the “Humanity’s Last Exam” benchmark Claude slightly edges Gemini (53.1% vs 51.4%). Code ability is within 1% across models, while Gemini leads terminal programming by about 3 points.

Direct Visual Comparison

Google posted a side‑by‑side video showing faster response and higher quality from 3.1 Pro versus 3.0 on the same tasks.

Practical Demos

City Planner Application

The model generated a full city‑planning app handling complex terrain, road and infrastructure layout, traffic flow simulation, and 3D visualization, integrating constraints and multi‑objective optimization without requiring the user to write code.

3‑D Flocking Simulation

A demo of starling flocking produced code implementing boids algorithms, hand‑gesture interaction, dynamic audio, and 3‑D rendering, illustrating “creative programming” beyond business logic.

Real‑time Code Verification

During a code‑review session, Gemini 3.1 Pro detected uncertainty about a SQL query and automatically launched a PostgreSQL container to test it, demonstrating a “try‑and‑verify” reasoning style.

Long‑Context Performance

While the model supports up to 1 M tokens input and 64 k tokens output, measured performance drops sharply: 128 k context yields 84.9% accuracy, but 1 M context falls to 26.3%, indicating current limits for ultra‑long texts.

Availability

Gemini 3.1 Pro is already live on Gemini App (free trial), Google AI Studio, Vertex AI, and NotebookLM (Pro/Ultra subscription). API pricing matches the previous Gemini 3 Pro.

References

Google DeepMind model page: https://deepmind.google/models/gemini/pro/

Sundar Pichai announcement: https://x.com/sundarpichai/status/2024516418855981298

Google tweet: https://x.com/google/status/2024519455389192204

3.0 vs 3.1 comparison video: https://x.com/google/status/2024598510499426433

Gemini API documentation: https://ai.google.dev/gemini-api/docs/