Benchmarks & tests
Published scores for the models we track, side by side. Every number comes from a model card, an official announcement or a public leaderboard — click a score to see its source.
Read this first
Labs run benchmarks with different settings (reasoning effort, tool access, number of attempts), so a 2-point gap rarely means much. Use scores to shortlist, then test the models on your own tasks. Empty cells mean no published score we could verify.
Coding, agents, reasoning and knowledge tests.Language & multimodal models
Blind-preference arena ratings (Elo) for generative media.Image, video & voice models
| Model | Image Arena Elo (AA) | Video Arena Elo (AA) | Voice Arena Elo (AA) |
|---|---|---|---|
| Nano Banana 2.1 |
1173 (rank 4) | — | — |
| FLUX 3 Image Black Forest Labs |
1115 (rank 13) | — | — |
| Eleven v4 ElevenLabs |
— | — | 1328±17 (rank #2, behind Eleven v4 Turbo) |
| Eleven v4 Turbo ElevenLabs |
— | — | 1330±18 (rank #1) |
| GPT Image 2.5 (Flare and Sunburst) OpenAI |
1217 (rank 1) | — | — |
| Gemini Omni Flash 1.1 |
— | 1132 (rank 8) | — |
| Kling 3.0 (Video 3.0 and Omni) Kling AI (Kuaishou) |
— | 1020 (rank 16) | — |
| FLUX.2 ([max], [pro], [flex], [dev], [klein]) Black Forest Labs |
1027 (rank 27) | — | — |
| Nano Banana Pro (Gemini 3 Pro Image) |
1115 (rank 12) | — | — |
| Veo 3.1 (Standard, Fast, Lite) |
— | 968 (rank 20) | — |
GlossaryWhat the tests measure
Terminal-Bench
Tasks the model must complete by running commands in a real shell. Measures agentic, tool-using work.
GPQA Diamond
Graduate-level multiple-choice questions in biology, physics and chemistry, written so that web search does not help. Measures expert reasoning.
Humanity's Last Exam
Very hard questions across many subjects written by domain experts. Low scores are normal; it tests the frontier.
SWE-bench
Real GitHub issues; the model must write a patch that makes the project's tests pass. Measures practical coding and debugging.
OSWorld
The model operates a real desktop computer to complete tasks. Measures computer-use agents.
GDPval
Realistic work tasks from many professions, graded against expert output. Measures economically useful knowledge work.
MMLU-Pro
Harder multiple-choice knowledge test across many academic subjects. Older MMLU is close to saturated.
LiveCodeBench
Fresh competitive-programming problems collected over time, which limits training-data contamination.
Arena Elo
Blind side-by-side comparisons where people vote for the better output (text, image, video or voice); Elo ratings move over time.