Google Releases Gemini 3.1 Pro: 77.1% ARC-AGI-2, Doubled Reasoning Performance

Google DeepMind Launches Gemini 3.1 Pro

Google DeepMind officially launched Gemini 3.1 Pro on February 19, 2026, marking a significant advancement in multimodal reasoning and agentic AI capabilities. The model is the latest in the Gemini 3 series.

Benchmark Results

The model achieves 77.1% on ARC-AGI-2—more than doubling its predecessor Gemini 3 Pro's score of 31.1%. On SWE-Bench Verified, it scores 80.6%, and on GPQA Diamond (graduate-level science reasoning), it reaches 94.3%. Terminal-Bench 2.0 performance stands at 68.5%, and Humanity's Last Exam (with tools) scores 51.4%.

Technical Specifications

The model supports a 1 million token input context window, enabling processing of over 1,500 pages of text or entire code repositories in a single prompt. Maximum output is 65,536 tokens. Gemini 3.1 Pro natively handles text, images, audio, video, and code in unified multimodal workflows.

Key Improvements

A new granular thinking mode lets users select reasoning depth from minimal to high, balancing speed and accuracy. Hallucination rates dropped significantly from 88% to 50% on the AA-Omniscience benchmark. Agentic reliability has been improved for autonomous multi-step workflows.

Pricing and Access

Pricing remains unchanged: $2/1M tokens (input up to 200K), $4/1M (input beyond 200K), $12/1M (output up to 200K), $18/1M (output beyond 200K). Available through Google AI Studio, Vertex AI, Gemini API, NotebookLM, and Microsoft Foundry. Deep Think mode is available to Google AI Ultra subscribers.

Source: Google DeepMind — Gemini 3.1 Pro Model Card