At the end of March last year, Google released Gemini 2.5 Pro, officially calling it "the most intelligent model" to date.
Every company makes such claims, but this time Gemini 2.5 Pro's benchmark scores are genuinely impressive:
Math and Science Reasoning
- AIME 2024: 92.0%
- AIME 2025: 86.7%
- Humanity's Last Exam: 18.8% (highest among models not using tools)
The last test was created by hundreds of domain experts specifically to probe the "boundaries of human knowledge," and achieving nearly 20% accuracy is quite remarkable.
Code and Multimodal
- LiveCodeBench: Ranked first in competitive programming
- MMMU: 84.0% (multimodal reasoning)
- WebDevArena Elo: 1443 (ranked first on the leaderboard)
Context Capabilities
It comes standard with a 1 million token context window (expandable to 2 million), supporting multimodal input including text, audio, images, and video. An entire code repository can be fed in directly for understanding without needing to be sliced.
Gemini 2.5 Pro is a "thinking model" — it performs internal reasoning before responding. This design approach is in the same vein as OpenAI's o-series and DeepSeek R1, but Google's implementation performs more evenly across multimodal scenarios.
However, one issue has been repeatedly raised in actual use: response speed. Deep reasoning means longer wait times, which diminishes the experience in scenarios requiring quick interaction. This is a trade-off common to all thinking models.
On LMArena, its Elo score is 1470, up 24 points from before its release, maintaining its lead on the leaderboard. Google has been criticized in the AI space for a long time, and this time it has responded with product strength.
Sources: CocoLoop, VentureBeat report, Google official release