Epoch AI conducted a third-party evaluation of the publicly available version of Google's Gemini 2.5 Deep Think model, assessing its mathematical reasoning and performance on benchmarks like FrontierMath.
- The model achieved state-of-the-art scores on FrontierMath Tiers 1–3 (29%) and Tier 4 (10%), demonstrating improved background knowledge and computational precision.
- Two professional mathematicians characterized the model as a generally helpful research assistant, broadly comparable to other top-tier models.
- While the model earned a bronze medal-equivalent score on the 2025 International Mathematical Olympiad (IMO), it struggled with problems requiring high creativity or intricate proofs.
- The evaluation noted that Deep Think relies more on conceptual approaches for geometry and occasionally provided incorrect citations of mathematical literature.
The report highlights the model's strengths in executing complex computations while noting its limitations in creative problem-solving and factual accuracy regarding sources.