Gemini 3.1 Pro: 2x 3.0 on ARC-AGI 2

2026-02-19 · Source: AINews · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Advanced, extended

Summary

Google has released Gemini 3.1 Pro, a new frontier model available as a developer preview via API/AI Studio and Vertex AI, and integrated into the Gemini app and NotebookLM. This model is positioned as the scaled-down core intelligence of Gemini 3 Deep Think. It demonstrates significant reasoning improvements, achieving an ARC-AGI-2 score of 77.1%, more than double that of Gemini 3 Pro. Additionally, it shows strong performance in coding and agentic-tool benchmarks, such as SWE-Bench Verified at 80.6%, and improved hallucination behavior. Independent evaluations largely corroborate its top-tier performance and competitive cost-to-intelligence ratio, with pricing remaining at $2/$12 per 1M input/output tokens for contexts up to 200k. Community reactions are mixed, with excitement over practical gains in areas like SVG design and web UI, alongside skepticism regarding benchmark-targeting and rollout inconsistencies.

Key takeaway

For CTOs and VPs of Engineering evaluating new LLMs for product integration, Gemini 3.1 Pro offers compelling benchmark improvements in reasoning and coding at a competitive price point. Your teams should pilot its capabilities for agentic workflows and creative design tasks, but be mindful of reported rollout inconsistencies and ensure robust testing for real-world agentic performance, as its GDPval scores still lag some competitors. Prioritize clear integration paths and consistent availability over raw benchmark numbers.

Key insights

Gemini 3.1 Pro advances reasoning and coding capabilities, achieving top-tier benchmarks while maintaining competitive pricing.

Principles

• Benchmark performance does not always equate to real-world agentic task leadership.
• Cost-to-intelligence ratio is a critical factor for model adoption.
• Effective product rollout and integration are as vital as model capabilities.

Method

Google scaled down the core intelligence of Gemini 3 Deep Think to create Gemini 3.1 Pro, focusing on improving reasoning, coding, and agentic-tool benchmarks, and reducing hallucination rates for practical product use.

In practice

• Explore Gemini 3.1 Pro for SVG design, web UI generation, and code quality tasks.
• Utilize its 1M context window and 64k max output for complex reasoning tasks.
• Consider its $2/$12 per 1M token pricing for cost-effective deployments.

Topics

• Gemini 3.1 Pro
• AI Benchmarking
• Agentic AI
• LLM Deployment
• Model Performance

Code references

• KittenML/KittenTTS
• karmaniverous/n8n-nodes-openclaw
• evaleval/every_eval_ever
• microsoft/DirectML
• alarmclock-kisser/OnnxBpmScanner

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• The handshake refusal heard around the AI world — Intense competition and rapid advancements define the current landscape of leading AI development.
• Gemini 3.1 Pro: A Hands-On Test of Google’s Newest AI — Gemini 3.1 Pro significantly advances reasoning, context handling, and agentic capabilities while reducing hallucinations.
• 😺 Google’s sharpest brain yet? 🧠 — Google's Gemini 3.1 Pro now leads in overall intelligence and factual accuracy, offering a cost-effective, advanced AI model.
• Gemini 3.1 Pro in 9 mins! — Gemini 3.1 Pro offers strong performance in coding and agentic tasks at a significantly lower cost than competitors.
• Gemini 3.0 Pro GA WILL BE Google's Greatest Model Ever! Most Powerful AI EVER! (Early Test) — Gemini 3.0 GA shows strong agentic task performance and advanced code generation, particularly for SVG and functional UI.
• 🗞️ Google releases Gemini 3 Deep Think, tops ARC-AGI 2 Benchmark With 84.6% — Advanced AI models are achieving human-level performance in complex reasoning and code generation through enhanced architectural and operational methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AINews.