Gemini Deep Think

Google’s “Deep Think” reasoning model is now publicly available after helping DeepMind reach gold-medal-level performance at the International Mathematical Olympiad (IMO), a benchmark long treated as a stress test for high-school-level mathematical reasoning. The key takeaway isn’t just that the system solved IMO problems—it did so using a different approach than earlier DeepMind attempts, and it can take dramatically longer to produce results than standard chat-style models. That tradeoff—stronger reasoning at the cost of speed—runs through every example.

IMO, held since 1959, asks teams of students to solve six extremely difficult problems across areas like algebra, geometry, and number theory. Each problem is scored out of seven for a maximum of 42 points. DeepMind previously achieved a silver-equivalent result, but the comparison wasn’t considered fully “fair” because the AI system used far more time than human contestants were allowed. This year’s breakthrough is framed as more comparable: rather than relying on specialized math-checking systems such as AlphaProof or AlphaGeometry, DeepMind used an advanced Gemini variant called “Deep Think,” feeding it the IMO questions directly.

The transcript also places the IMO results in a broader AI arms race. OpenAI announced that an experimental reasoning LLM reached gold-medal-level performance, releasing proofs for the five problems it solved. Rumors at ICML in Vancouver suggested Google may also have scored gold, with timing around announcements reportedly designed to avoid overshadowing human winners. Both companies publicly released only the solutions for the problems they got right, raising the question of whether the remaining problems were too awkward to disclose.

To gauge Deep Think’s practical behavior, the transcript describes hands-on testing in which an IMO problem was pasted into the Gemini interface without using Lean or other formal proof tooling. The most striking observation is latency: after several minutes, no answer appeared, and even when “thinking tokens” began to stream, the final solution arrived much later—about 16 minutes in one run. Another benchmark problem from the AIME 2025 dataset showed a similar pattern: the model produced intermediate summaries that already pointed to the correct numeric answer (204), yet continued “thinking” for several additional minutes before returning the final response.

Outside math, Deep Think was tested on a prompt to generate a “salatai” scene using colorful voxels and 3D code. The result was treated as a success: the model produced usable Three.js code, rendered a navigable 3D environment, and demonstrated detailed understanding of voxel-based structure.

A final experiment involved an Angry Birds-style setup. With Deep Think enabled, the model detected missing runtime dependencies (it flagged the absence of “pygame”), iterated on the level design, and improved physics enough to score when hitting pigs. Still, the workflow required longer waits than a non-Deep-Think model, leading to skepticism about using Deep Think as a default coding engine.

Overall, Deep Think is portrayed as impressive for hard reasoning tasks—especially math and logic—but its long “parallel thinking” process makes it less practical when speed and cost matter. The transcript ends on the idea that future model development will hinge on balancing intelligence against latency and expense.