🏆 AI Benchmarks Index
What each major benchmark measures and who leads it — every score served from our own tracking pipeline and refreshed daily. JSON is available at /api/benchmarks.
Graduate-level science questions designed to be Google-proof — tests deep reasoning.
- 1GPT-6.2 Astraopenai96.8%
- 2Gemini 4 Hydrogengoogle96.8%
- 3Gemini 4 Carbongoogle96.5%
- 4Gemini 4 Argongoogle96.4%
- 5Claude Fable 5.2anthropic96.0%
- 6GPT-6 Astraopenai95.8%
- 7Claude Sonnet 5.5anthropic95.6%
- 8GPT-6.1 Solopenai95.4%
- 9Gemini 3.8 Flashgoogle95.4%
- 10Gemini 3.7 Flashgoogle94.8%
Real GitHub issues an agent must fix end-to-end — the standard for coding agents.
- 1Claude Opus 4.7 (Fast)anthropic83.5%
- 2Claude Opus 4.7anthropic83.5%
- 3GPT-5.5openai80.6%
- 4Gemini 3.5 Flashgoogle79.3%
- 5Claude Opus 4.6anthropic78.7%
- 6GLM 5.2z-ai78.7%
- 7DeepSeek V4 Prodeepseek77.6%
- 8Qwen3.7 Maxqwen77.3%
- 9GPT-5.4openai76.9%
- 10Claude Opus 4.5anthropic76.7%
The hardest tier of competition math problems from the MATH benchmark.
- 1GPT-5openai98.1%
- 2GPT-5 Miniopenai97.8%
- 3o4 Miniopenai97.8%
- 4o3openai97.8%
- 5Claude Sonnet 4.5anthropic97.7%
- 6Qwen3 Maxqwen97.1%
- 7R1deepseek96.6%
- 8o3 Miniopenai96.5%
- 9Claude Haiku 4.5anthropic96.4%
- 10Gemini 2.5 Progoogle95.9%
Olympiad-style mock AIME exam (2024–25) — elite high-school competition math.
- 1Qwen3.8 Max (0902)qwen100.0%
- 2Claude Opus 5.5anthropic100.0%
- 3GPT-6.2 Astraopenai100.0%
- 4GPT-5.6 Solopenai100.0%
- 5GPT-5.5 Proopenai100.0%
- 6Gemini 4 Hydrogengoogle100.0%
- 7Claude Fable 5anthropic100.0%
- 8Gemini 4 Carbongoogle100.0%
- 9GPT-5.5openai100.0%
- 10Claude Sonnet 5.5anthropic100.0%
Research-level math problems written by mathematicians — the hardest math benchmark.
- 1GPT-5.5 Proopenai52.4%
- 2GPT-5.5openai51.7%
- 3GPT-5.4 Proopenai50.0%
- 4GPT-5.4openai47.6%
- 5Claude Opus 4.8 (Fast)anthropic47.2%
- 6Claude Opus 4.8anthropic47.2%
- 7Claude Opus 4.7anthropic43.8%
- 8Claude Opus 4.7 (Fast)anthropic43.8%
- 9GPT-5.2openai40.7%
- 10Claude Opus 4.6anthropic40.7%
Short factual questions measuring hallucination resistance — can it just be right?
- 1Claude Fable 5.2anthropic94.0%
- 2GPT-6.2 Astraopenai78.5%
- 3Gemini 4 Carbongoogle77.7%
- 4Gemini 4 Argongoogle76.0%
- 5GPT-6 Astraopenai75.6%
- 6GPT-6.1 Solopenai73.9%
- 7Claude Opus 5.5anthropic72.2%
- 8Gemini 4 Hydrogengoogle70.9%
- 9Claude Fable 5.1anthropic70.8%
- 10Claude Fable 5anthropic70.7%
Human preference Elo from blind head-to-head chat battles — the de-facto vibes benchmark.
- 1claude-opus-5-highanthropic1502
- 2Claude Opus 4.6anthropic1497
- 3Claude Fable 5anthropic1493
- 4MiMo-V2.6-Proxiaomi1492
- 5claude-fable-5-highanthropic1492
- 6Qwen3.8 Maxqwen1486
- 7Claude Opus 4.7anthropic1483
- 8Gemini 3.5 Flashgoogle1482
- 9Gemini 3.1 Pro Previewgoogle1480
- 10Muse Spark 1.1meta1479
Human preference Elo for text-to-image generation battles.
- 1gpt-image-2.5-sunburstopenai1429
- 2GPT-Image-2.6 Sunburst (Medium)openai1410
- 3gpt-image-2.5-flareopenai1400
- 4MAI-Image-2.7microsoft-ai1387
- 5gpt-image-2 (medium)openai1382
- 6GPT-Image-2.6 Flare (Medium)openai1375
- 7grok-imagine-image-2.0 (canvas)xai1334
- 8mai-image-2.6microsoft-ai1331
- 9Nano Banana 2.1google1328
- 10Grok Imagine Image 2.0xai1318
Human preference Elo for image-to-video generation battles.
- 1Gemini Omni 1.2 Flashgoogle1503
- 2dreamina-seedance-2.5-720pbytedance1494
- 3minimax-h3minimax1493
- 4gemini-omni-1.1-flashgoogle1487
- 5wan3.0alibaba1482
- 6dreamina-seedance-2.0-720pbytedance1480
- 7SkyReels V4skyreel1475
- 8grok-imagine-video-1.5-preview-720pxai1466
- 9gemini-omni-flashgoogle1463
- 10hidream-o1-video-1.0hidream1458
Score data: Epoch AI Benchmarking Hub (CC-BY) and LMArena (CC-BY), ingested daily and matched to our model catalog. Rankings are never for sale — see the methodology and full model leaderboard.