- Humanity's Last Exam (no tools) 12.1
- GPQA diamond single attempt (pass@1) 78.3
- AIME 2025 single attempt 78
- AIME 2024 single attempt 88
- LiveCodeBench v5 single attempt (10/1/2024-2/1/2025) 63.5
- Aider Polyglot whole 51.1 / diff-fenced 44.2 51.1
- SimpleQA 29.7
- MMMU single attempt 76.7
- Vibe-Eval (Reka) 62
- MRCR 128k (average) (1M pointwise 66.3) 84.6
- Global MMLU (Lite) 88.4
Gemini 2.5 Flash Apr 2025
Gemini 2.5 Flash places in the top three on 5 benchmarks.
33
Benchmarks with a score
0
Ranked first, out of five or more models
5
In the top three
Google DeepMind Apr 2025 Gemini 2.5 Flash
Apr 2025 snapshot Closed weights
140.8
#122 of 256
Epoch Capabilities Index
Self-Reported Index 136.3 #216 of 404
- Humanity's Last Exam 12.1% #26 /47
- LiveCodeBench v5 63.5% #20 /47
- VideoReasonBench 27.4% #2 /14
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LongBench v2 | 62.1% | #2 of 33 | Official leaderboard | |
| Vectara Hallucination Leaderboard (HHEM) | 1.30% | #10 of 148 | Official leaderboard | |
| Arena-Hard v2.0 | 68.6% | #4 of 23 | Official leaderboard | |
| Global MMLU Lite | 88.4% | #3 of 16 | Lab model card | |
| MRCR (Michelangelo, original) | 84.6% | #2 of 9 | Lab model card | |
| SimpleQA | 29.7% | #25 of 60 | Lab model card | |
| Humanity's Last Exam | 12.1% | #26 of 47 | Official leaderboard | |
| Fiction.LiveBench | 47.2% | #43 of 57 | Fiction.live leaderboard, via Epoch AI |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AIME 2024 | 88% | #24 of 88 | Lab model card | |
| ChemBench | 64.1% | #7 of 24 | Official leaderboard | |
| GPQA Diamond | 78.3% | #130 of 302 | Lab model card | |
| OTIS Mock AIME 2024-2025 | 73.1% | #105 of 212 | Epoch AI run | |
| MATH-500 | 91.8% | #19 of 38 | Official leaderboard | |
| Putnam-like | 84% | #3 of 6 | Official leaderboard | |
| AIME 2025 | 78% | #94 of 143 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench v5 | 63.5% | #20 of 47 | Lab model card | |
| LiveCodeBench | 81.8% | #13 of 21 | Official leaderboard | |
| LiveCodeBench v6 | 60.6% | #13 of 21 | Official leaderboard | |
| Aider Polyglot | 51.1% | #20 of 30 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| lmgame-Bench | 24.6 | #13 of 23 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LegalBench | 83.8% | #47 of 138 | Official leaderboard | |
| TaxEval v2 | 71.2% | #77 of 129 | Official leaderboard | |
| CorpFin v2 | 54.2% | #88 of 119 | Official leaderboard | |
| MortgageTax | 58.4% | #67 of 87 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| VideoReasonBench | 27.4% | #2 of 14 | Official leaderboard | |
| VISTA | 47% | #18 of 54 | Official leaderboard | |
| MMMU | 76.7% | #35 of 103 | Lab model card | |
| GeoBench | 73% | #23 of 48 | Official leaderboard | |
| Vibe-Eval | 62% | #5 of 9 | Lab model card | |
| Point-Bench | 47% | #10 of 17 | Official leaderboard | |
| VPCT | 38% | #18 of 27 | VPCT leaderboard, via Epoch AI |
Its lab's numbers put it at 136.3, 4.4 points below its ECI
The Self-Reported Index fits one score to the 11 capability results Google DeepMind reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 134.4 to 138.3. How we compute it.
- Global MMLU Lite 88.4
- AIME 2024 88
- MRCR (Michelangelo, original) 84.6
- GPQA Diamond 78.3
- AIME 2025 78
- MMMU 76.7
- LiveCodeBench v5 63.5
- Vibe-Eval 62
- Aider Polyglot 51.1
- SimpleQA 29.7
- Humanity's Last Exam 12.1
Its lab's numbers next to 3 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| Aider Polyglot | 51.1% Aider Polyglot whole 51.1 / diff-fenced 44.2 | 47.1% Leaderboard, aider, diff edits | +4.0 pts |
| LiveCodeBench v5 | 63.5% LiveCodeBench v5 single attempt (10/1/2024-2/1/2025) | 62.2% Leaderboard | +1.3 pts |
| Humanity's Last Exam | 12.1% Humanity's Last Exam (no tools) | 12.1% Leaderboard | Same |
Google DeepMind's models over time
Google DeepMind's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 5 of them aren't in the launch post.
Gemini 2.5 Flash Preview Model Card Apr 29, 2025
Capabilities
- Humanity's Last Exam 12.1
- GPQA Diamond 78.3
- AIME 2025 78
- AIME 2024 88
- LiveCodeBench v5 63.5
- Aider Polyglot 51.1
- SimpleQA 29.7
- MMMU 76.7
- Vibe-Eval 62
- MRCR (Michelangelo, original) 84.6
- Global MMLU Lite 88.4