Gemini 2.0 Flash Thinking crushes math and science benchmarks

Google's Gemini 2.0 Flash Thinking is getting a fast upgrade.

Demis Hassabis shared on X that the latest update to the model just scored 73.3% on the AIME math benchmark and 74.2% on the GPQA Diamond science test. That is a noticeable jump from its initial rollout last December.

Code execution and fewer logic slip-ups

Beyond raw test scores, the update brings native code execution alongside a massive 1M token context window.

It also tackles a well-known quirk in reasoning models by reducing thought-answer contradictions. Users often see reasoning models work through a problem correctly in their hidden chain of thought, only to spit out an incorrect final answer at the very end.

Hassabis noted that DeepMind has been developing these types of planning systems for over a decade, tracing back to programs like AlphaGo, and that merging those concepts with frontier foundation models is proving effective.

Why it matters

Fixing thought-answer discrepancies is a very practical win. If a model cannot reliably translate its own internal reasoning into the final response, high benchmark scores mean little in real-world workflows.

Combining that fix with code execution and a 1M token window on a Flash-tier model is a smart move. We think it shows Google is aiming to make multi-step reasoning fast, cheap, and long-context-ready, putting pressure on competitors whose best reasoning tools still feel sluggish and expensive.