Editorial illustration for Google Gemini 3.1 Pro doubles reasoning performance in benchmark
Google Gemini 3.1 Pro doubles reasoning performance in...
Google's latest AI model doesn't just edge ahead—it vaults. On the ARC-AGI-2 benchmark, a test of novel abstract thought, Gemini 3.1 Pro scored 77.1%. That more than doubles its predecessor's 31.1% and decisively beats Anthropic's best. The same story holds on Humanity's Last Exam, a rigorous academic test, where it hit 44.4% without tools.
On ARC-AGI-2, a benchmark that evaluates a model's ability to solve novel abstract reasoning patterns, 3.1 Pro scored 77.1% — more than double the 31.1% achieved by Gemini 3 Pro and substantially ahead of Anthropic's Sonnet 4.6 (58.3%) and Opus 4.6 (68.8%). This result also eclipses OpenAI's GPT-5.2 (52.9%).
VentureBeat called it a "Deep Think Mini," highlighting its adjustable reasoning depth. That feature—the ability to dial up computational effort for a tough problem—could define its real-world utility far more than any benchmark score. These numbers reset the entire conversation.
For months, the race seemed a game of minor leads. A doubling on core reasoning is a spike. It creates unambiguous distance, placing direct pressure on rivals to answer with a leap of their own.
Related Reading
Browse all Google news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks
From Archives NVIDIA Co-Design Boosts Sarvam AI Inference, Cuts TTFT Below One Second
From Archives Anthropic's mid-tier model offers 30‑minute ChatGPT crash course, 100+ prompts
From Archives Google unveils Gemini 3.1 Pro, hits 94.3% GPQA Diamond and coding Elo 2
Research & Benchmarks
From Archives Gemini 3.1 Pro released as upgraded Gemini 3 Deep Think for complex tasks
Common Questions Answered
How much did Gemini 3.1 Pro improve on the ARC-AGI-2 benchmark compared to its predecessor?
Gemini 3.1 Pro scored 77.1% on the ARC-AGI-2 benchmark, more than doubling its predecessor's score of 31.1%. This represents a significant leap in performance on a test designed to measure novel abstract thought capabilities.
What is the 'Deep Think Mini' feature in Gemini 3.1 Pro and why is it important?
The 'Deep Think Mini' feature allows users to adjust the reasoning depth of the model, enabling them to dial up computational effort when tackling difficult problems. According to VentureBeat, this adjustable reasoning capability could define the model's real-world utility more than benchmark scores alone.
How did Gemini 3.1 Pro perform on Humanity's Last Exam benchmark?
Gemini 3.1 Pro achieved a score of 44.4% on Humanity's Last Exam, which is described as a rigorous academic test. This performance was achieved without the use of external tools, demonstrating strong reasoning capabilities on complex academic problems.
How does Gemini 3.1 Pro's reasoning performance compare to Anthropic's best model?
Gemini 3.1 Pro decisively beats Anthropic's best model on the ARC-AGI-2 benchmark, scoring 77.1% compared to Anthropic's lower performance on the same abstract reasoning test. This performance gap represents a significant competitive advantage in core reasoning capabilities.
Further Reading
- Gemini 3.1 Pro: A smarter model for your most complex tasks — Google Blog
- Google Releases Gemini 3.1 Pro and Lyria 3 — Case Western Reserve University
- Google’s new Gemini Pro model has record benchmark scores — again — TechCrunch
- Gemini 3.1 Pro Leads Most Benchmarks But Trails Claude Opus 4.6 ... — Trending Topics
