- AIME2024 (Math) 73.3
- GPQA Diamond (Science) 74.2
- MMMU (Multimodal reasoning) 75.4
Gemini 2.0 Flash Thinking Jan 2025
Gemini 2.0 Flash Thinking ranks first on EMMA.
39
Benchmarks with a score
1
Ranked first, out of five or more models
4
In the top three
Google DeepMind Jan 2025 Gemini 2.0 Flash Thinking
Jan 2025 snapshot Closed weights
133.8
#148 of 256
Epoch Capabilities Index
- Humanity's Last Exam 6.56% #38 /47
- ZeroBench 5.00% #51 /76
- GPQA Diamond 57.1% #140 /229
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 | |
|---|---|---|---|---|
| Vectara Hallucination Leaderboard (HHEM) | 1.80% | #20 of 148 | Official leaderboard | |
| LLM Confabulation Leaderboard | 12.4% | #10 of 58 | Official leaderboard | |
| SEAL Multilingual | 1,108 | #4 of 19 | Official leaderboard | |
| LongBench v2 | 56% | #9 of 33 | Official leaderboard | |
| AidanBench | 1531.0 | #5 of 13 | Official leaderboard | |
| Fiction.LiveBench | 52.8% | #39 of 57 | Fiction.live leaderboard, via Epoch AI | |
| Humanity's Last Exam | 6.56% | #38 of 47 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| U-MATH | 83.6% | #2 of 30 | Official leaderboard | |
| PolyMath | 37.1% | #14 of 38 | Official leaderboard | |
| 3SAT-Search | 62.6% | #5 of 11 | Official leaderboard | |
| TSP-Opt | 30.1% | #6 of 11 | Official leaderboard | |
| PutnamBench | 0.15% | #7 of 12 | Official leaderboard | |
| OTIS Mock AIME 2024-2025 | 57.8% | #127 of 212 | Epoch AI run | |
| USAMO 2025 | 4.17% | #6 of 10 | Official leaderboard | |
| GPQA Diamond | 57.1% | #140 of 229 | Epoch AI run | |
| AIME 2024 | 73.3% | #55 of 88 | Lab model card | |
| GSM8K-Platinum | 99.2% | #5 of 8 | Official leaderboard | |
| PHYBench | 18.2% | #19 of 28 | Official leaderboard | |
| MATH-500 | 84.6% | #30 of 38 | Official leaderboard | |
| SimpleBench | 30.7% | #84 of 106 | Official leaderboard | |
| HMMT 2025 | 35.8% | #42 of 53 | Official leaderboard | |
| AIME 2025 | 53.3% | #44 of 53 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SEAL Coding | 1,108 | #6 of 26 | Official leaderboard | |
| BigCodeBench-Hard | 27% | #32 of 96 | Official leaderboard | |
| Aider Polyglot | 18.2% | #37 of 46 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ToolComp | 63.2% | #11 of 32 | Official leaderboard | |
| Berkeley Function Calling Leaderboard (BFCL) | 58.7% | #26 of 67 | Official leaderboard | |
| Step Race | 2 | #30 of 59 | Official leaderboard | |
| Elimination Game | 3 | #49 of 56 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FinanceReasoning | 85.1% | #3 of 6 | Official leaderboard | |
| TaxEval v2 | 69.8% | #89 of 129 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MMVU | 69.1% | #2 of 28 | Official leaderboard | |
| EMMA | 42.5% | #1 of 7 | Official leaderboard | |
| EmoBench-M | 60.6% | #5 of 15 | Official leaderboard | |
| MMMU | 75.4% | #40 of 103 | Lab model card | |
| VISTA | 45.5% | #24 of 54 | Official leaderboard | |
| SpaCE-10 | 42.2% | #18 of 36 | Official leaderboard | |
| GeoBench | 70% | #29 of 48 | Official leaderboard | |
| ZeroBench | 5.00% | #51 of 76 | Official leaderboard |
Its lab's number next to a leaderboard
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 |
|---|---|---|---|
| GPQA Diamond | 74.2% GPQA Diamond (Science) | 57.1% Epoch AI run | +17 pts |
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.