Google Expands Gemini Deep Think Access with New API

Google has implemented a two-pronged release for the enhanced Gemini Deep Think. Subscribers to the Google AI Ultra plan can now access the updated reasoning mode directly within the Gemini application. Concurrently, the company has launched an early access program allowing select researchers, engineers, and enterprise clients to integrate Deep Think into their own tools and workflows via the Gemini API. This marks the first time the advanced reasoning mode has been made available through a developer interface.

According to the company, Deep Think was developed in collaboration with scientists and is designed for complex problems that may lack a single correct answer or involve incomplete data. The API access is intended to facilitate the integration of this capability into existing research and engineering processes, enabling automated tasks and custom application development.

To demonstrate the capabilities of the upgraded model, Google cited performance across several academic and competitive programming benchmarks. The company claims the following results:

• A score of 48.4% on Humanity’s Last Exam, a test designed to challenge frontier AI models, without the use of external tools.
• An 84.6% score on the ARC-AGI-2 benchmark.
• An Elo rating of 3455 on Codeforces, a platform for competitive programming.
• Gold-medal-level performance in the International Math Olympiad 2025.
• Gold-medal-level results on the written portions of the 2025 International Physics Olympiad and the Chemistry Olympiad.
• A score of 50.5% on CMT-Benchmark, which Google describes as an evaluation for advanced theoretical physics.

Google stated the release is part of a broader strategy to shift from abstract reasoning evaluations toward practical applications in research and engineering. By making Deep Think available through an API, the company aims to embed its advanced reasoning capabilities into professional workflows where it can be used to interpret complex data sets and model physical systems using code. The company also noted it is developing specialized agents that utilize Deep Think for high-level mathematics research.

The announcement did not include details on the future pricing structure for API access to Gemini Deep Think following the early access period. The specific criteria for selection into the early access program and the number of available slots have not been disclosed. Furthermore, Google provided no timeline for the public release or operational details of the specialized research agents it mentioned are in development.

The immediate next step is the rollout of the Gemini API early access program, for which Google is currently accepting expressions of interest from scientists, engineers, and enterprises. For individual users, the feature is now live within the Gemini app for those on the appropriate subscription tier. The move is expected to allow developers and researchers to begin building and testing applications that leverage the model’s advanced reasoning for specialized domains like theoretical physics and competitive mathematics.

Based on the announcement, interested parties can take several actions. Google AI Ultra subscribers can explore the updated Deep Think mode within the Gemini app. Researchers, engineers, and enterprise users interested in programmatic access can apply for the early access program through Google’s official channels. Developers can begin evaluating how advanced reasoning models could be integrated into their existing software and data analysis pipelines.