Model Leaderboard
Which AI model is actually best — real benchmarks and real usage, scheduled to refresh daily. Not vibes.
LLM Stats: last updated Oct 12, 2026
Free preview
Every table below is open to everyone except Free Right Now, which is Premium/Founding only. Premium also unlocks BREAF's own written take on why a model earned its spot.
See PremiumHow the BREAF Score works
The daily calculation uses up to 36 benchmarks selected across all ten skill areas. Each benchmark score is converted to a percentile against today's current models — not a raw number, since different benchmarks use wildly different scales. Those percentiles are grouped into ten skill areas (Reasoning, Coding, Agents, and so on) using llm-stats.com's own categorization of each benchmark, then combined into one 0-100 BREAF Score.
A model needs both enough evidence and evidence spread across enough skill areas to be trusted at face value — a handful of large, well-attested benchmarks (say, Humanity's Last Exam) can count as much as many small ones, but scoring well only on a narrow slice of skills doesn't. Whichever is missing pulls the score toward the middle of the scale, deliberately, so a model that's only ever been run where it wins can't out-rank one that's been tested broadly and honestly, warts included. "Tested on" shows how many named benchmarks actually fed a model's score.
Click any row to see the specific named benchmarks — MMLU-Pro, GPQA, SWE-Bench Verified, AIME 2025, and more — behind its category scores. A dash in any column means that model hasn't been tested there, not a zero — no lab runs every benchmark.
A "Preview" tag means llm-stats.com has no public release date for that model yet — it's ranked on its real scores like everything else, just not something you can use today.
About 96% of the underlying numbers are reported by the model's own maker, not independently checked — llm-stats.com publishes them as-is, and so do we. Scores within about a point of each other are close enough to call a tie, not a real ranking. This composite is BREAF's own calculation over llm-stats.com's published benchmark data, not a number llm-stats.com itself publishes.
Source: llm-stats.com.
10 newest models — best available score
| # | Model | BREAF Score | Tested on | Reasoning | Math | Coding | Agents | Tools | Search | Vision | Long Ctx | Writing | General | Price in/M | Price out/M | Speed | License |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropic | 70.0 | 9 bench | 88% | — | 96% | 89% | 76% | — | 100% | — | — | 88% | $4.00 | $20.00 | — | Proprietary |
| 2 | GPT-6 AstraOpenAI | 69.8 | 8 bench | 87% | — | 85% | 84% | 82% | 93% | 87% | — | — | 95% | $10.00 | $50.00 | — | Proprietary |
| 3 | Muse Spark 1.3Meta | 68.6 | 5 bench | 88% | — | 94% | 85% | 85% | — | 60% | 100% | — | 78% | $0.10 | $0.20 | — | Proprietary |
| 4 | GPT-6.1 SolOpenAI | 67.9 | 5 bench | 89% | — | 85% | 87% | — | — | 91% | — | — | 91% | $2.00 | $10.00 | — | Proprietary |
| 5 | Claude Fable 5.1Anthropic | 66.4 | 4 bench | 73% | 100% | 77% | 68% | 60% | — | 97% | — | — | 93% | $10.00 | $50.00 | — | Proprietary |
| 6 | Gemini 4 ArgonGoogle | 65.2 | 7 bench | 72% | — | 91% | 78% | 72% | — | 81% | 90% | — | 67% | — | — | — | Proprietary |
| 7 | Kimi K3Moonshot AI | 63.5 | 15 bench | 72% | 75% | 80% | 71% | 71% | 87% | 75% | — | — | 64% | $2.85 | $14.25 | — | Kimi K3 License |
| 8 | Claude Opus 5Anthropic | 63.1 | 7 bench | 72% | 93% | 72% | 65% | 49% | 80% | 85% | — | — | 73% | $5.00 | $25.00 | — | Proprietary |
| 9 | Claude Sonnet 5.5Anthropic | 62.7 | 8 bench | 83% | — | 87% | 79% | 82% | — | — | — | — | 50% | $2.00 | $10.00 | 42 tok/s | Proprietary |
| 10 | 61.2 | 10 bench | 87% | 93% | 80% | 58% | — | 60% | 77% | 25% | — | 65% | — | — | — | Proprietary | |
| 11 | GPT-5.6 SolOpenAI | 60.7 | 14 bench | 68% | — | 71% | 63% | 59% | 73% | 72% | 73% | — | 76% | $5.00 | $30.00 | — | Proprietary |
| 12 | Claude Mythos 5.1Anthropic | 60.3 | 1 bench | 91% | — | 91% | 91% | 91% | — | — | — | — | — | — | — | — | Proprietary |
| 13 | Ling 3.1 FlashInclusionAI | 60.0 | 5 bench | 75% | — | 70% | 74% | 75% | — | — | — | — | — | — | — | — | Unknown |
| 14 | GLM-5.3Zhipu AI | 59.4 | 10 bench | 68% | 85% | 59% | 58% | 65% | — | 85% | — | — | — | $1.20 | $4.00 | — | GLM-5.3 License |
| 15 | GPT-6 SolOpenAI | 59.4 | 4 bench | 80% | — | 60% | 68% | 95% | — | 53% | — | — | 53% | $2.00 | $10.00 | — | Proprietary |
| 16 | MiMo-V2.6-ProXiaomi | 58.4 | 7 bench | 63% | — | 72% | 67% | 63% | — | 67% | — | — | 67% | $0.44 | $0.87 | — | MIT |
| 17 | Atria Dawn PreviewShanghai AI Laboratory | 58.4 | 6 bench | 67% | — | 53% | 68% | 65% | 100% | — | — | — | — | — | — | — | MIT |
| 18 | DeepSeek-V4-Pro-0813DeepSeek | 57.9 | 6 bench | 62% | 75% | 69% | 58% | 59% | — | 75% | — | — | — | $1.30 | $2.60 | — | MIT |
| 19 | Sakana NamazuSakana AI | 57.7 | 2 bench | 87% | 100% | — | — | — | — | — | — | — | 87% | $0.95 | $4.00 | — | Proprietary |
| 20 | Mistral Large 4Mistral AI | 57.5 | 5 bench | 75% | 100% | 46% | 49% | 66% | — | — | — | — | — | $0.68 | $2.09 | — | Proprietary |
| 21 | Ember-1Fireworks AI | 56.6 | 3 bench | 66% | — | 75% | 69% | 47% | — | — | — | — | — | $3.00 | $15.00 | — | Proprietary |
| 22 | Qwen3.8 MaxAlibaba Cloud / Qwen Team | 56.5 | 15 bench | 65% | 35% | 58% | 61% | 62% | — | 64% | 65% | — | 83% | $1.65 | $4.95 | — | Qwen3.8-Max License |
| 23 | Muse Spark 1.1Meta | 55.1 | 8 bench | 67% | 80% | 46% | 54% | 76% | — | 59% | — | — | 39% | $1.25 | $4.25 | — | Proprietary |
| 24 | Gemini 3.8 FlashGoogle | 54.5 | 6 bench | 51% | — | 66% | 62% | 59% | — | 47% | 86% | — | 40% | $0.75 | $3.75 | — | Proprietary |
| 25 | Hy4 previewTencent | 54.0 | 13 bench | 63% | — | 56% | 52% | 56% | — | — | — | — | 65% | — | — | — | Apache 2.0 |
| 26 | DeepSeek-V4.1-FlashDeepSeek | 53.4 | 9 bench | 61% | 30% | 71% | 71% | 72% | — | 30% | — | — | 45% | $0.22 | $0.66 | — | MIT |
| 27 | Seed 2.1 ProByteDance | 52.3 | 16 bench | 53% | 62% | 45% | 43% | 44% | 53% | 52% | 43% | — | 89% | — | — | — | Proprietary |
| 28 | Claude Haiku 5.5Anthropic | 52.2 | 6 bench | 61% | — | 72% | 62% | 59% | — | — | — | — | 13% | $0.10 | $0.50 | — | Proprietary |
| 29 | Grok 4.7xAI | 51.3 | 3 bench | 55% | — | 52% | 52% | 55% | — | — | — | — | — | $2.00 | $6.00 | — | Proprietary |
| 30 | GLM-5.3-FlashZhipu AI | 51.3 | 8 bench | 60% | 55% | 49% | 55% | 65% | — | 60% | — | — | 25% | $0.15 | $0.50 | — | MIT |
| 31 | Gemini 3.8 Flash CyberGoogle | 51.2 | 1 bench | — | — | 67% | 67% | — | — | — | — | — | — | — | — | — | Proprietary |
| 32 | Gemini 3.7 FlashGoogle | 50.2 | 9 bench | 50% | — | 52% | 40% | 44% | — | 49% | 79% | — | 49% | $0.75 | $3.75 | — | Proprietary |
| 33 | MiMo-V2.6-FlashXiaomi | 50.1 | 7 bench | 51% | — | 61% | 54% | 51% | — | 39% | — | — | 39% | $0.14 | $0.28 | — | MIT |
| 34 | Qwen3.8 FlashAlibaba Cloud / Qwen Team | 49.8 | 13 bench | 65% | 36% | 47% | 50% | 77% | — | 39% | 7% | — | 61% | $0.15 | $0.47 | — | Proprietary |
| 35 | Qwen3.8-Flash-NextAlibaba Cloud / Qwen Team | 49.8 | 13 bench | 65% | 36% | 47% | 50% | 77% | — | 39% | 7% | — | 61% | — | — | — | Qwen Community License 1.0 |
| 36 | Gemini 3.5 Flash CyberGoogle | 49.8 | 1 bench | — | — | 47% | 47% | — | — | — | — | — | — | — | — | — | Proprietary |
| 37 | Muse Spark 1.2Meta | 49.3 | 2 bench | 55% | — | 43% | 43% | 55% | — | — | — | — | — | $0.10 | $0.20 | — | Proprietary |
| 38 | Grok 4.5xAI | 49.2 | 6 bench | 52% | — | 38% | 38% | 34% | — | — | — | — | 82% | $2.00 | $6.00 | 80 tok/s | Proprietary |
| 39 | GPT-5.6 TerraOpenAI | 48.4 | 14 bench | 47% | — | 54% | 47% | 42% | 67% | 37% | 27% | — | 52% | $2.00 | $12.00 | — | Proprietary |
| 40 | Ling 3.0 Flash FinInclusionAI | 48.2 | 1 bench | 14% | — | — | 14% | — | — | — | — | — | — | $0.06 | $0.18 | — | Unknown |
| 41 | DeepSeek-V4-Flash-Vision-ExpDeepSeek | 47.2 | 3 bench | 37% | — | 61% | 37% | 37% | — | — | — | — | — | $0.44 | $1.32 | — | Unknown |
| 42 | GPT-6 LunaOpenAI | 47.2 | 4 bench | 49% | — | 36% | 45% | 71% | — | 33% | — | — | 33% | $0.10 | $0.50 | — | Proprietary |
| 43 | 47.1 | 3 bench | 45% | 61% | — | — | — | — | 40% | — | — | 23% | — | — | — | Proprietary | |
| 44 | Gemini 3.6 FlashGoogle | 45.7 | 6 bench | 43% | — | 28% | 36% | 39% | — | 71% | 29% | — | 56% | $1.50 | $7.50 | — | Proprietary |
| 45 | Seed 2.1 TurboByteDance | 45.2 | 15 bench | 42% | 48% | 36% | 33% | 30% | 47% | 37% | 29% | — | 74% | — | — | — | Proprietary |
| 46 | 45.1 | 1 bench | 0% | 0% | — | — | — | — | — | — | — | 0% | — | — | — | Apache 2.0 | |
| 47 | Qwen3.8-27BAlibaba Cloud / Qwen Team | 45.0 | 11 bench | 45% | 19% | 30% | 42% | 45% | — | 52% | — | — | 58% | $0.40 | $3.00 | — | Apache 2.0 |
| 48 | Grok 4.6xAI | 44.6 | 3 bench | 46% | — | 33% | 44% | 23% | — | — | — | — | — | $2.00 | $6.00 | — | Proprietary |
| 49 | Claude Sonnet 5Anthropic | 44.1 | 13 bench | 40% | 70% | 41% | 29% | 14% | 40% | 57% | — | — | 29% | $2.00 | $10.00 | 42 tok/s | Proprietary |
| 50 | GLM-5.2Zhipu AI | 43.8 | 10 bench | 35% | 50% | 33% | 28% | 29% | — | 50% | — | — | 50% | $0.75 | $2.40 | — | MIT |
Synced Oct 12 · Text/Image/Video/TTS/STT: Data by LLM Stats. Benchmarks: Source: Artificial Analysis, via OpenRouter (openrouter.ai/rankings), licensed CC BY 4.0. Most Used: Source: OpenRouter (openrouter.ai/rankings), licensed CC BY 4.0. Free Right Now: OpenRouter (licensed CC BY 4.0), NVIDIA NIM, and SambaNova Cloud — each provider's own public model catalog. Text scores are scheduled daily; other sources are scheduled every 6 hours.