🏆 AI Benchmarks Index

Graduate-level science questions designed to be Google-proof — tests deep reasoning.

1. 1Gemini 3.7 Flashgoogle94.8%
2. 2GPT-5.4 Proopenai94.6%
3. 3Gemini 3.6 Flashgoogle94.1%
4. 4Grok 4.6x-ai94.0%
5. 5GPT-5.5openai94.0%
6. 6GPT-5.5 Proopenai93.9%
7. 7Claude Opus 5anthropic93.9%
8. 8GPT-5.6 Solopenai93.5%
9. 9Grok 4.5x-ai93.4%
10. 10GPT-5.6 Terraopenai93.3%

Real GitHub issues an agent must fix end-to-end — the standard for coding agents.

1. 1Claude Opus 4.7anthropic83.5%
2. 2Claude Opus 4.7 (Fast)anthropic83.5%
3. 3GPT-5.5openai80.6%
4. 4Gemini 3.5 Flashgoogle79.3%
5. 5Claude Opus 4.6anthropic78.7%
6. 6GLM 5.2z-ai78.7%
7. 7DeepSeek V4 Prodeepseek77.6%
8. 8Qwen3.7 Maxqwen77.3%
9. 9GPT-5.4openai76.9%
10. 10Kimi K2.6moonshotai76.7%

The hardest tier of competition math problems from the MATH benchmark.

1. 1GPT-5openai98.1%
2. 2GPT-5 Miniopenai97.8%
3. 3o4 Miniopenai97.8%
4. 4o3openai97.8%
5. 5Claude Sonnet 4.5anthropic97.7%
6. 6Qwen3 Maxqwen97.1%
7. 7R1deepseek96.6%
8. 8o3 Miniopenai96.5%
9. 9Claude Haiku 4.5anthropic96.4%
10. 10Gemini 2.5 Progoogle95.9%

Olympiad-style mock AIME exam (2024–25) — elite high-school competition math.

1. 1Claude Fable 5anthropic100.0%
2. 2GPT-5.5 Proopenai100.0%
3. 3GPT-5.5openai100.0%
4. 4GPT-5.6 Solopenai100.0%
5. 5GPT-5.6 Terraopenai99.7%
6. 6Grok 4.6x-ai99.2%
7. 7Claude Opus 5anthropic98.9%
8. 8DeepSeek V4 Pro 0813deepseek98.6%
9. 9GPT-5.6 Lunaopenai98.3%
10. 10Claude Opus 4.8anthropic98.3%

Research-level math problems written by mathematicians — the hardest math benchmark.

1. 1GPT-5.5 Proopenai52.4%
2. 2GPT-5.5openai51.7%
3. 3GPT-5.4 Proopenai50.0%
4. 4GPT-5.4openai47.6%
5. 5Claude Opus 4.8anthropic47.2%
6. 6Claude Opus 4.8 (Fast)anthropic47.2%
7. 7Claude Opus 4.7 (Fast)anthropic43.8%
8. 8Claude Opus 4.7anthropic43.8%
9. 9GPT-5.2openai40.7%
10. 10Claude Opus 4.6anthropic40.7%

Short factual questions measuring hallucination resistance — can it just be right?

1. 1GPT-5.6 Solopenai71.6%
2. 2Gemini 3.7 Flashgoogle71.2%
3. 3Gemini 3.6 Flashgoogle68.7%
4. 4Gemini 3.5 Flashgoogle68.4%
5. 5Claude Fable 5anthropic68.3%
6. 6Qwen3 Maxqwen67.5%
7. 7GPT-5.5 Proopenai64.5%
8. 8GPT-5.5openai63.1%
9. 9Qwen3.7 Maxqwen58.5%
10. 10DeepSeek V4 Prodeepseek57.0%

Human preference Elo from blind head-to-head chat battles — the de-facto vibes benchmark.

1. 1Claude Opus 4.6anthropic1497
2. 2Claude Fable 5anthropic1495
3. 3Claude Opus 4.7anthropic1483
4. 4Gemini 3.5 Flashgoogle1482
5. 5Qwen3.8 Maxqwen1482
6. 6Gemini 3.1 Pro Previewgoogle1480
7. 7Gemini 3.6 Flashgoogle1479
8. 8Muse Spark 1.1meta1478
9. 9Kimi K3moonshotai1473
10. 10ernie-5.1baidu1468

Human preference Elo for text-to-image generation battles.

1. 1gpt-image-2 (medium)openai1381
2. 2reve-2.1reve1302
3. 3muse-imagemeta1283
4. 4gemini-3.1-flash-image-preview (nano-banana-2) [web-search]google1270
5. 5reve-2.0reve1270
6. 6gemini-3.1-flash-image (nano-banana-2) [web-search]google1263
7. 7qwen-image-3.0-proalibaba1258
8. 8seedream-5.0-probytedance1257
9. 9mai-image-2.5microsoft-ai1256
10. 10gemini-3.1-flash-lite-image (nano-banana-2-lite)google1251

Human preference Elo for image-to-video generation battles.

1. 1minimax-h3minimax1489
2. 2dreamina-seedance-2.5-720pbytedance1484
3. 3dreamina-seedance-2.0-720pbytedance1478
4. 4grok-imagine-video-1.5-preview-720pxai1466
5. 5gemini-omni-flashgoogle1462
6. 6grok-imagine-video-1.5-720pxai1460
7. 7flux-3-video-20260811bfl1449
8. 8happyhorse-1.0aorizon1442
9. 9wan2.7-i2vwan1428
10. 10grok-imagine-video-720pxai1415