Artificial intelligence · February 2026

Gemini 3 Deep Think

Extended reasoning passes 84% on ARC-AGI-2.

Google's upgraded Deep Think mode scored 84.6% on ARC-AGI-2, verified by the ARC Prize, 48.4% on Humanity's Last Exam without tools, and a Codeforces rating of 3455. A benchmark designed to resist language models fell within a year of its release.

See it in the living tree Wikipedia Source

Arcs above join it to what it stood on, arcs below to what it led to. The bar places this stretch on the tree's whole timeline.

What it built on 2

Gemini 3 2025
Test-time compute scaling 2024

Related measurements

The dashed line marks when it first worked.

Best AI score on GPQA Diamond Latest: 95.8 % correct (2026) Source: Epoch AI, Capabilities & benchmarking (benchmark_data.zip, gpqa_diamond.csv); Anthropic: Claude 3 public launch (March 4, 2024)

Training compute of notable AI models Source: Epoch AI, Notable AI Models (CC BY 4.0)