Google expands Gemini Deep Think for research, APIs

Google has released an upgraded version of Gemini 3 Deep Think, a reasoning mode aimed at scientific, research, and engineering work. It has also widened access beyond the Gemini app by offering early access through the Gemini API to a selected group of users.

The updated Deep Think mode is available in the Gemini app for Google AI Ultra subscribers. Researchers, engineers, and enterprises can also apply for early access to test Deep Think via the Gemini API, the first time this mode has been offered through Google's developer interface.

Deep Think is a specialised reasoning mode within the Gemini 3 model family. Google says it was developed in close collaboration with scientists and researchers, and is designed for research settings where problems may not have a single correct answer and data can be incomplete.

Benchmark claims

Google cited a set of academic and competitive benchmarks to describe improvements in the upgraded Deep Think mode. It reported 48.4% on Humanity's Last Exam without tools, which it described as a test designed to probe the limits of frontier AI models.

It also reported 84.6% on ARC-AGI-2, a result Google said was verified by the ARC Prize Foundation, and an Elo rating of 3455 on Codeforces, a competitive programming platform.

Google also said the updated Deep Think reached gold-medal-level performance on the International Math Olympiad 2025. It linked the upgrade to earlier work in which specialised versions of Deep Think tackled difficult reasoning challenges in maths and programming competitions.

Science domains

Beyond mathematics and coding, Google said the updated mode performs strongly in chemistry and physics. It reported gold-medal-level results on the written sections of the 2025 International Physics Olympiad and the Chemistry Olympiad.

Google also highlighted theoretical physics performance, reporting a score of 50.5% on CMT-Benchmark, which it described as an advanced theoretical physics evaluation.

Engineering use

Google is positioning Deep Think for research and engineering workflows as well as academic-style evaluations. It said the mode can help interpret complex data and model physical systems through code.

Making Deep Think available through the Gemini API changes how it can be integrated into existing tools and processes. API access is typically used to build applications, connect models to internal systems, and run automated tasks at scale. Google said access will initially be limited to select users through the early access programme.

The move ties an advanced reasoning mode to both a consumer subscription tier and a developer channel. The Gemini app remains the primary entry point for individual subscribers, while the API targets professional and organisational deployment.

Google framed the upgrade as part of a broader effort to move from abstract reasoning tests toward practical research and engineering work. It also said it has been developing specialised agents that use Deep Think for research-level mathematics exploration, without detailing how those agents operate or where they are being used.

Access to the updated Deep Think mode is now live for Google AI Ultra subscribers in the Gemini app. The Gemini API early access programme is accepting expressions of interest from scientists, engineers, and enterprises.