The History of AI

Gemini 3.1 Pro

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models

Lineage
Neurons and networks
Kind
Implementation

Who

Google DeepMind

What happened

Google releases Gemini 3.1 Pro, an upgrade of Gemini 3 Pro focused on reasoning. It scores 77.1 percent on the ARC-AGI-2 benchmark, ahead of Claude Opus 4.6 and GPT-5.2.

Why it mattered

Better reasoning: 77.1 percent on ARC-AGI-2.

The full story

On February 19, 2026, three months after Gemini 3 Pro, Google released Gemini 3.1 Pro as a preview. It builds on Gemini 3 Pro and is designed for complex reasoning tasks with lots of data and varied inputs: text, audio, images, video, and entire codebases.

The architecture is still a mixture-of-experts Transformer, which activates only part of its parameters for each response, with up to a million input tokens and 64,000 output tokens.

The headline figure was on ARC-AGI-2, where it scored 77.1 percent, far ahead of GPT-5.2 and also ahead of Claude Opus 4.6, with all three models in their most compute-intensive modes. Google also claimed record scores on Terminal-Bench 2.0, for agentic coding, and on MCP Atlas.

According to Google, the underlying improvement had appeared a week earlier in Gemini 3 Deep Think, a mode for scientific tasks, with which early users had found a flaw in a peer-reviewed mathematics paper.

The model reached developer tools, the Gemini app, and NotebookLM. Two weeks later OpenAI released GPT-5.4.

Sources

How to cite

“Gemini 3.1 Pro.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gemini-3-1-pro

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline