Gemini 3.1 Pro Boosts Reasoning Ability by 148% – What’s New?

What is ARC‑AGI‑2?

ARC‑AGI‑2 is a widely‑recognized hard reasoning benchmark that tests a model’s ability to infer unseen logical patterns rather than memorise answers, thus reflecting true problem‑understanding.

Gemini 3.1 Pro’s Leap

Google announced Gemini 3.1 Pro on February 19. The model achieved a 77.1% score on ARC‑AGI‑2, up from roughly 31.1% for Gemini 3 Pro—a 46‑point increase representing a 148% improvement. By contrast, Google’s Gemini 3 Deep Think scores 84.6% on the same test, only 7.5 points higher.

Where the Upgrade Is Strong

Reasoning Ability

Google positions the model for complex tasks, moving from simple answer generation to handling problems that cannot be solved in a single sentence. For example, asking the model to “create a website” now yields a complete animated SVG instead of raw HTML, and requesting an ISS telemetry dashboard produces a functional visual instrument panel.

Agentic Capabilities

The announcement repeatedly mentions “agentic workflows.” Unlike earlier chat‑style models, Gemini 3.1 Pro can plan and execute ongoing tasks, such as monitoring a web‑server log and notifying on anomalies, demonstrating early steps toward autonomous agents.

Product Matrix

Google’s lineup now covers three dimensions:

Gemini 3 Flash – optimized for speed, suitable for low‑latency chat bots.

Gemini 3.1 Pro – focused on reasoning, ideal for complex analysis, code generation, system design.

Gemini 3 Deep Think – built for deep reasoning tasks such as physics‑level problem solving.

This tiered approach lets users pick capabilities without overpaying for unused features.

How It Stacks Up Against Claude and GPT

Reasoning

On ARC‑AGI‑2, Gemini 3.1 Pro’s 77.1% places it among the industry leaders, while Deep Think’s 84.6% is the current public high. Claude Opus 4.6 and GPT‑5.2 have not disclosed scores, so direct dominance cannot be claimed.

Programming

Claude Opus 4.6 scores 81.42% on SWE‑bench Verified (real GitHub issue fixing) and 65.4% on Terminal‑Bench 2.0 (complex terminal tasks), leading the field. Gemini 3.1 Pro has not released comparable data, indicating programming is not its primary focus.

Knowledge Work

Claude also leads on GDPval‑AA, outperforming GPT‑5.2 by about 144 Elo points, showing strength in finance, law and other professional domains.

Practical Features Beyond Benchmarks

SVG Animation Generation

Gemini 3.1 Pro can turn textual prompts like “create a tech‑style loading animation with blue‑purple gradient” into ready‑to‑use SVG code, saving designers and front‑end developers considerable time.

Multimodal Abilities

The model retains Gemini’s multimodal support for images, video and audio, enabling deeper analysis such as diagnosing posture issues from a video clip.

Pricing and Access

Google has not announced a price; the model is currently available for free within Google AI Studio and the Gemini App under usage limits.

Caveats and Limitations

Preview Status

The release is a preview, meaning the current version may differ from the final product in performance and stability.

Self‑Reported Benchmarks

All scores come from Google’s own testing, which may be influenced by test design, environment and dataset selection. Independent third‑party evaluations are needed for objective validation.

Benchmark Scope

ARC‑AGI‑2 measures abstract reasoning only; everyday tasks like code writing, content creation or translation may not reflect the same level of improvement.

Recommendations

Developers should experiment with complex logical reasoning, SVG/code generation, and multimodal tasks in the free preview.

Ordinary users are advised to wait for the stable release and third‑party test results.

When selecting a model for production, avoid locking into a single provider; design a flexible architecture that can swap model back‑ends as the landscape evolves.

Final Thoughts

The 77.1% ARC‑AGI‑2 score demonstrates that AI reasoning competition is accelerating, but the race is far from settled. Different models excel in different domains, so users should choose tools that best match their specific needs.

Related Links

Gemini 3.1 Pro official announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/

Google AI Studio (free trial): https://aistudio.google.com/

Gemini 3 Deep Think announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

Claude Opus 4.6 announcement: https://www.anthropic.com/news/claude-opus-4-6