Google upgrades Gemini 3 Deep Think to V2

Google shipped a major upgrade to Gemini 3 Deep Think, its specialised high-effort reasoning mode, developed in what the company described as close partnership with scientists and researchers on real research problems. Where the previous Deep Think had been strongest on mathematics and programming, the update extended that performance into broader scientific domains including chemistry and physics.

Google reported scores of 48.4% on Humanity’s Last Exam without external tools, 84.6% on ARC-AGI-2 (verified by the ARC Prize Foundation), and an Elo rating of 3455 on Codeforces, alongside gold-medal-level performance on the 2025 International Mathematical, Physics and Chemistry Olympiads and 50.5% on the CMT-Benchmark for theoretical physics. The model was made available in the Gemini app for Google AI Ultra subscribers, with API access opened to a select group of researchers, engineers and enterprises through an early-access programme.

Google’s own account of Gemini 3.1 Pro, released the following week, described that model’s core reasoning as drawing on the intelligence that had debuted with this Deep Think upgrade — tying the two releases together as a staged rollout of the same underlying reasoning improvements, first to a high-effort specialised mode and then to the general-purpose flagship.