Gemini 3 Deep Think (major upgrade)

2026-02-12 · Publisher : Google · Domain : Artificial Intelligence

Google updates Gemini 3 Deep Think, its specialized reasoning mode for science, research and engineering, with new records on ARC-AGI-2 and Humanity's Last Exam.

What changes

Upgrade designed with scientists and researchers, for problems with incomplete data and no single answer.
Better results in mathematics, competitive programming, physics and chemistry (gold-medal level on the written parts of the 2025 physics and chemistry olympiads).
Applied use: going from a sketch to a 3D-printable file, reviewing mathematical proofs.
For the first time, access to Deep Think via the Gemini API, on request in an early access program.

### Specifications - Availability: Gemini app (Google AI Ultra subscribers) and Gemini API in early access. - Context window, maximum output, knowledge cutoff and API identifier: not disclosed in the post. - Reasoning mode of the Gemini 3 family, not a separate model listed in the public pricing. ### Pricing Pricing not disclosed at launch (API access on request). ### Benchmarks | Benchmark (scores published by the vendor) | Gemini 3 Deep Think | |---|---| | Humanity's Last Exam (no tools) | 48.4% | | ARC-AGI-2 (ARC Prize verified) | 84.6% | | Codeforces (Elo) | 3455 | | CMT-Benchmark (theoretical physics) | 50.5% | ### Key takeaways A long reasoning mode reserved for Ultra subscribers and a few companies, useful for assessing the limits of scientific reasoning; API access remains restricted.

Gemini 3 Deep Think (major upgrade)

What changes

Links