AI models
How the labs evolved
Each lab's record per benchmark over time, and price per token when each model was listed on OpenRouter.
Record per lab: GPQA Diamond ?what it is
- Google 96.4%
- OpenAI 95.8%
- Anthropic 95.6%
- Moonshot AI 94.2%
- xAI 94%
- DeepSeek 93.2%
- ByteDance 93.1%
- Alibaba 92.7%
- Zhipu AI (Z.ai) 91.9%
- MiniMax 90.9%
- Mistral AI 90.2%
- Meta 89.8%
- NVIDIA 85.4%
- Microsoft 56.1%
- Ai2 (Allen Institute) 46.3%
- 01.AI 32%
Source: Epoch AI, CC BY 4.0, latest version we checked. The 90% to 100% band marks the saturation zone.
Chart data table
| Lab | Model | % | Dates in sources |
|---|---|---|---|
| Anthropic | Claude 2 | 34.7% | 2023 |
| Anthropic | Claude 3 Sonnet | 40.6% | 2024 |
| Anthropic | Claude 3 Opus | 47.2% | 2024 |
| Anthropic | Claude 3.5 Sonnet | 54% | 2024 |
| Anthropic | Claude 3.5 Sonnet (October 2024) | 55.3% | 2024 |
| Anthropic | Claude 3.7 Sonnet | 66% | 2025 |
| Anthropic | Claude 3.7 Sonnet | 76.8% | 2025 |
| Anthropic | Claude 3.7 Sonnet | 79.7% | 2025 |
| Anthropic | Claude Sonnet 4.5 | 81.7% | 2025 |
| Anthropic | Claude Sonnet 4.5 | 82.3% | 2025 |
| Anthropic | Claude Opus 4.5 | 85.5% | 2025 |
| Anthropic | Claude Opus 4.5 | 86% | 2025 |
| Anthropic | Claude Opus 4.6 | 88.4% | 2026 |
| Anthropic | Claude Opus 4.6 | 88.8% | 2026 |
| Anthropic | Claude Opus 4.6 | 90.5% | 2026 |
| Anthropic | Claude Opus 4.8 | 91% | 2026 |
| Anthropic | Claude Opus 5 | 92.9% | 2026 |
| Anthropic | Claude Opus 5 | 93.9% | 2026 |
| Anthropic | Claude Sonnet 5.5 | 95.6% | 2026 |
| Moonshot AI | Kimi K2 Thinking | 84.2% | 2025 |
| Moonshot AI | Kimi K2.5 | 87.6% | 2026 |
| Moonshot AI | Kimi K2.6 | 90.8% | 2026 |
| Moonshot AI | Kimi K3 | 91.9% | 2026 |
| Moonshot AI | Kimi K3 | 93.1% | 2026 |
| Moonshot AI | Kimi K3.1 Pro | 94.2% | 2026 |
| ByteDance | Doubao-Seed-1.6 | 77.6% | 2025 |
| ByteDance | Doubao-Seed-2.0 Pro | 88.2% | 2026 |
| ByteDance | Doubao Seed 2.1 Pro | 89.3% | 2026 |
| ByteDance | Doubao Seed 2.1 Turbo | 89.6% | 2026 |
| ByteDance | Doubao Seed Evolving (Nov 2026) | 93.1% | 2026 |
| Mistral AI | Mistral 7B | 13.2% | 2023 |
| Mistral AI | Mixtral 8x7B | 29.8% | 2023 |
| Mistral AI | Mixtral 8x7B | 30.6% | 2023 |
| Mistral AI | Mistral Large | 38.8% | 2024 |
| Mistral AI | Mistral Large 2 | 49% | 2024 |
| Mistral AI | Mistral Large 2 | 51.3% | 2024 |
| Mistral AI | Mistral Medium 3 | 59.5% | 2025 |
| Mistral AI | Mistral Large 4 | 88.1% | 2026 |
| Mistral AI | Mistral Large 4 | 90.2% | 2026 |
| OpenAI | GPT-4 (Mar 2023) | 35.7% | 2023 |
| OpenAI | GPT-4 Turbo (Nov 2023) | 42.4% | 2023 |
| OpenAI | GPT-4 Turbo (Apr 2024) | 46.6% | 2024 |
| OpenAI | GPT-4o | 48.9% | 2024 |
| OpenAI | GPT-4o | 49.2% | 2024 |
| OpenAI | o1-preview | 50.3% | 2024 |
| OpenAI | o1-mini | 59.5% | 2024 |
| OpenAI | o1-mini | 62.4% | 2024 |
| OpenAI | o1 | 74.2% | 2024 |
| OpenAI | o1 | 75.8% | 2024 |
| OpenAI | o1 | 76.8% | 2024 |
| OpenAI | o3-mini | 77% | 2025 |
| OpenAI | o4-mini | 77.8% | 2025 |
| OpenAI | o4-mini | 79.6% | 2025 |
| OpenAI | o3 | 79.8% | 2025 |
| OpenAI | o3 | 80.8% | 2025 |
| OpenAI | o3 | 81.8% | 2025 |
| OpenAI | GPT-5 | 85.4% | 2025 |
| OpenAI | GPT-5 | 86.2% | 2025 |
| OpenAI | GPT-5.1 | 87.6% | 2025 |
| OpenAI | GPT-5.2 | 87.9% | 2025 |
| OpenAI | GPT-5.2 | 88.2% | 2025 |
| OpenAI | GPT-5.2 | 91.4% | 2025 |
| OpenAI | GPT-5.4 | 93.3% | 2026 |
| OpenAI | GPT-5.4 Pro | 94.6% | 2026 |
| OpenAI | GPT-6 Astra | 95.8% | 2026 |
| DeepSeek | DeepSeek LLM 67B | 24.6% | 2023 |
| DeepSeek | DeepSeek-V3 | 56.5% | 2024 |
| DeepSeek | DeepSeek-R1-Distill-Qwen-32B | 64.1% | 2025 |
| DeepSeek | DeepSeek-R1 | 71.7% | 2025 |
| DeepSeek | DeepSeek-R1 (May 2025) | 76.3% | 2025 |
| DeepSeek | DeepSeek-V3.2 | 83.4% | 2025 |
| DeepSeek | DeepSeek-V4-Pro | 89.6% | 2026 |
| DeepSeek | DeepSeek-V4-Pro | 90.9% | 2026 |
| DeepSeek | DeepSeek V4 Flash 0731 | 91% | 2026 |
| DeepSeek | DeepSeek V4 Pro 0813 | 91.7% | 2026 |
| DeepSeek | DeepSeek V4.1 Flash 1017 | 93.2% | 2026 |
| Meta | Llama 2-70B | 26.3% | 2023 |
| Meta | Llama 3-70B | 40.6% | 2024 |
| Meta | Llama 3.1-70B | 44.2% | 2024 |
| Meta | Llama 3.1-405B | 50.9% | 2024 |
| Meta | Llama 4 Scout | 51.8% | 2025 |
| Meta | Llama 4 Maverick | 67% | 2025 |
| Meta | Muse Spark | 89.8% | 2026 |
| Gemini 1.0 Pro | 34% | 2023 | |
| Gemini 1.5 Pro | 45.9% | 2024 | |
| Gemini 1.5 Flash | 47.3% | 2024 | |
| Gemini 1.5 Pro | 57.2% | 2024 | |
| Gemini 2.0 Flash | 64.1% | 2025 | |
| Gemini 2.0 Pro | 65.7% | 2025 | |
| Gemini 2.5 Pro (Mar 2025) | 83.8% | 2025 | |
| Gemini 2.5 Pro (Jun 2025) | 84.8% | 2025 | |
| Gemini 2.5 Pro (Jun 2025) | 85.3% | 2025 | |
| Gemini 3 Pro | 92.6% | 2025 | |
| Gemini 3.1 Pro | 94.1% | 2026 | |
| Gemini 3.1 Pro | 94.4% | 2026 | |
| Gemini 3.7 Flash | 94.8% | 2026 | |
| Gemini 3.8 Flash | 95.4% | 2026 | |
| Gemini 4 Argon | 96.4% | 2026 | |
| xAI | Grok-2 | 53.8% | 2024 |
| xAI | Grok-3 mini | 75.5% | 2025 |
| xAI | Grok 3 | 75.8% | 2025 |
| xAI | Grok-3 mini | 76.3% | 2025 |
| xAI | Grok 4 | 87% | 2025 |
| xAI | Grok 4.20 | 89.3% | 2026 |
| xAI | Grok 4.5 | 93.4% | 2026 |
| xAI | Grok 4.6 | 94% | 2026 |
| Alibaba | Qwen1.5-72B | 28.8% | 2024 |
| Alibaba | Qwen1.5-32B | 30.7% | 2024 |
| Alibaba | Qwen2-72B | 40.8% | 2024 |
| Alibaba | Qwen2.5-32B | 46.1% | 2024 |
| Alibaba | Qwen2.5-72B | 49.1% | 2024 |
| Alibaba | Qwen2.5-Max | 56.1% | 2025 |
| Alibaba | QwQ-32B | 65.3% | 2025 |
| Alibaba | QWQ-Plus | 65.4% | 2025 |
| Alibaba | Qwen3-32B | 65.7% | 2025 |
| Alibaba | Qwen3-235B-A22B | 70.7% | 2025 |
| Alibaba | Qwen3-235B-A22B (Jul 2025) | 80.1% | 2025 |
| Alibaba | Qwen3.5 397B-A17B | 85.9% | 2026 |
| Alibaba | Qwen3.5 397B-A17B | 86.4% | 2026 |
| Alibaba | Qwen 3.6 Plus | 88.4% | 2026 |
| Alibaba | Qwen3.7-Max | 90.9% | 2026 |
| Alibaba | Qwen 3.8 Max | 92.7% | 2026 |
| Zhipu AI (Z.ai) | GLM-4.7 | 83.3% | 2025 |
| Zhipu AI (Z.ai) | GLM-5 | 87.8% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.1 | 89.9% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 91.9% | 2026 |
| NVIDIA | Nemotron 3 Ultra | 85.4% | 2026 |
| MiniMax | MiniMax-M3 | 81.3% | 2026 |
| MiniMax | MiniMax-M3 | 90.9% | 2026 |
| Microsoft | WizardLM-2 8x22B | 43.4% | 2024 |
| Microsoft | Phi-4 | 56.1% | 2024 |
| 01.AI | Yi-34B | 14.7% | 2023 |
| 01.AI | Yi-1.5-34B | 32% | 2024 |
| Ai2 (Allen Institute) | Tulu 3 (Tülu 3) 70B | 46.3% | 2024 |
Record per lab: OTIS Mock AIME ?
- Anthropic 100%
- Moonshot AI 100%
- OpenAI 100%
- Google 100%
- Alibaba 100%
- Mistral AI 99.4%
- DeepSeek 99.4%
- Meta 99.4%
- ByteDance 99.2%
- xAI 99.2%
- Zhipu AI (Z.ai) 93.9%
- NVIDIA 86.7%
- MiniMax 71.1%
- Microsoft 13.8%
- Ai2 (Allen Institute) 4.4%
Source: Epoch AI, CC BY 4.0, latest version we checked. The 90% to 100% band marks the saturation zone.
Chart data table
| Lab | Model | % | Dates in sources |
|---|---|---|---|
| Anthropic | Claude 2 | 2.5% | 2023 |
| Anthropic | Claude 3 Opus | 4.7% | 2024 |
| Anthropic | Claude 3.5 Sonnet | 6.5% | 2024 |
| Anthropic | Claude 3.5 Sonnet (October 2024) | 8.5% | 2024 |
| Anthropic | Claude 3.7 Sonnet | 21.9% | 2025 |
| Anthropic | Claude 3.7 Sonnet | 46.7% | 2025 |
| Anthropic | Claude 3.7 Sonnet | 53.3% | 2025 |
| Anthropic | Claude 3.7 Sonnet | 57.8% | 2025 |
| Anthropic | Claude Opus 4 | 60% | 2025 |
| Anthropic | Claude Opus 4 | 64.4% | 2025 |
| Anthropic | Claude Sonnet 4 | 68.9% | 2025 |
| Anthropic | Claude Sonnet 4 | 71.1% | 2025 |
| Anthropic | Claude Sonnet 4.5 | 77.8% | 2025 |
| Anthropic | Claude Opus 4.5 | 81.7% | 2025 |
| Anthropic | Claude Opus 4.5 | 86.1% | 2025 |
| Anthropic | Claude Opus 4.6 | 91.1% | 2026 |
| Anthropic | Claude Opus 4.6 | 93.1% | 2026 |
| Anthropic | Claude Opus 4.6 | 94.4% | 2026 |
| Anthropic | Claude Opus 4.7 | 97.8% | 2026 |
| Anthropic | Claude Opus 4.8 | 98.3% | 2026 |
| Anthropic | Claude Fable 5 | 99.7% | 2026 |
| Anthropic | Claude Fable 5 | 100% | 2026 |
| Moonshot AI | Kimi K2 Thinking | 83.1% | 2025 |
| Moonshot AI | Kimi K2.5 | 92.2% | 2026 |
| Moonshot AI | Kimi K2.6 | 96.1% | 2026 |
| Moonshot AI | Kimi K3 | 97.2% | 2026 |
| Moonshot AI | Kimi K3.1 Pro | 100% | 2026 |
| ByteDance | Doubao-Seed-1.6 | 69.4% | 2025 |
| ByteDance | Doubao-Seed-2.0 Pro | 93.9% | 2026 |
| ByteDance | Doubao Seed 2.1 Pro | 95.6% | 2026 |
| ByteDance | Doubao Seed 2.1 Turbo | 96.1% | 2026 |
| ByteDance | Doubao Seed Evolving (Nov 2026) | 99.2% | 2026 |
| Mistral AI | Mistral Large | 1.9% | 2024 |
| Mistral AI | Mistral Large 2 | 8.5% | 2024 |
| Mistral AI | Mistral Medium 3 | 32.2% | 2025 |
| Mistral AI | Mistral Large 4 | 98.3% | 2026 |
| Mistral AI | Mistral Large 4 | 99.4% | 2026 |
| OpenAI | GPT-4 (Mar 2023) | 0.6% | 2023 |
| OpenAI | GPT-4 (Jun 2023) | 1.1% | 2023 |
| OpenAI | GPT-3.5 Turbo | 2.2% | 2024 |
| OpenAI | GPT-4 Turbo (Apr 2024) | 6.7% | 2024 |
| OpenAI | GPT-4o mini | 6.9% | 2024 |
| OpenAI | o1-preview | 31.1% | 2024 |
| OpenAI | o1-mini | 44.7% | 2024 |
| OpenAI | o1-mini | 46.9% | 2024 |
| OpenAI | o1 | 53.3% | 2024 |
| OpenAI | o1 | 73.3% | 2024 |
| OpenAI | o3-mini | 76.9% | 2025 |
| OpenAI | o4-mini | 81.7% | 2025 |
| OpenAI | o3 | 83.9% | 2025 |
| OpenAI | o3 | 84.4% | 2025 |
| OpenAI | gpt-oss-120b | 88.9% | 2025 |
| OpenAI | GPT-5 | 91.4% | 2025 |
| OpenAI | GPT-5.2 | 93.9% | 2025 |
| OpenAI | GPT-5.2 | 96.1% | 2025 |
| OpenAI | GPT-5.2 | 96.1% | 2025 |
| OpenAI | GPT-5.4 | 97.8% | 2026 |
| OpenAI | GPT-5.5 | 100% | 2026 |
| DeepSeek | DeepSeek LLM 67B | 0.8% | 2023 |
| DeepSeek | DeepSeek-V3 | 15.8% | 2024 |
| DeepSeek | DeepSeek-R1-Distill-Qwen-1.5B | 21.4% | 2025 |
| DeepSeek | DeepSeek-R1-Distill-Qwen-14B | 50.6% | 2025 |
| DeepSeek | DeepSeek-R1-Distill-Llama-70B | 51.4% | 2025 |
| DeepSeek | DeepSeek-R1 | 53.3% | 2025 |
| DeepSeek | DeepSeek-R1-Distill-Qwen-32B | 55.6% | 2025 |
| DeepSeek | DeepSeek-R1 (May 2025) | 66.4% | 2025 |
| DeepSeek | DeepSeek-V3.2 | 87.8% | 2025 |
| DeepSeek | DeepSeek-V4-Pro | 95.6% | 2026 |
| DeepSeek | DeepSeek-V4-Pro | 96.7% | 2026 |
| DeepSeek | DeepSeek V4 Pro 0813 | 98.6% | 2026 |
| DeepSeek | DeepSeek V4.1 Flash 1017 | 99.4% | 2026 |
| Meta | Llama 2-70B | 0% | 2023 |
| Meta | Llama 3-8B | 1.9% | 2024 |
| Meta | Llama 3-70B | 4.3% | 2024 |
| Meta | Llama 3.1-405B | 9.7% | 2024 |
| Meta | Llama 4 Maverick | 20.6% | 2025 |
| Meta | Muse Spark | 88.9% | 2026 |
| Meta | Muse Spark 1.3 | 99.2% | 2026 |
| Meta | Muse Wonder | 99.4% | 2026 |
| Gemini 1.0 Pro | 1.1% | 2023 | |
| Gemini 1.5 Pro | 6.8% | 2024 | |
| Gemini 1.5 Flash | 16.2% | 2024 | |
| Gemini 1.5 Pro | 23.1% | 2024 | |
| Gemini 2.0 Flash Thinking | 57.8% | 2025 | |
| Gemini 2.5 Flash | 73.1% | 2025 | |
| Gemini 2.5 Pro (Jun 2025) | 84.7% | 2025 | |
| Gemini 3 Pro | 91.4% | 2025 | |
| Gemini 3 Flash | 92.8% | 2025 | |
| Gemini 3 Flash | 95.6% | 2025 | |
| Gemini 3.1 Pro | 95.6% | 2026 | |
| Gemini 3.7 Flash | 97.2% | 2026 | |
| Gemini 3.8 Flash | 98.9% | 2026 | |
| Gemini 4 Argon | 100% | 2026 | |
| xAI | Grok-2 | 11.5% | 2024 |
| xAI | Grok 3 | 55.6% | 2025 |
| xAI | Grok-3 mini | 62.2% | 2025 |
| xAI | Grok-3 mini | 77.8% | 2025 |
| xAI | Grok 4 | 84% | 2025 |
| xAI | Grok 4.20 | 92.2% | 2026 |
| xAI | Grok 4.3 Beta | 93.3% | 2026 |
| xAI | Grok 4.5 | 97.8% | 2026 |
| xAI | Grok 4.6 | 99.2% | 2026 |
| Alibaba | Qwen2.5-32B | 7.4% | 2024 |
| Alibaba | Qwen2.5-72B | 8.1% | 2024 |
| Alibaba | Qwen2.5-Max | 16.1% | 2025 |
| Alibaba | Qwen Plus | 17.8% | 2025 |
| Alibaba | QwQ-32B | 59.2% | 2025 |
| Alibaba | Qwen3-30B-A3B | 62.8% | 2025 |
| Alibaba | Qwen3-14B | 66.4% | 2025 |
| Alibaba | Qwen3-32B | 66.9% | 2025 |
| Alibaba | Qwen3-235B-A22B (Jul 2025) | 86.7% | 2025 |
| Alibaba | Qwen3.5 397B-A17B | 88.9% | 2026 |
| Alibaba | Qwen 3.6 Plus | 93.3% | 2026 |
| Alibaba | Qwen3.7-Max | 95.6% | 2026 |
| Alibaba | Qwen 3.8 Max | 99.4% | 2026 |
| Alibaba | Qwen3.8 Max (0902) | 100% | 2026 |
| Zhipu AI (Z.ai) | GLM-4.7 | 83.3% | 2025 |
| Zhipu AI (Z.ai) | GLM-5.1 | 93.3% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.3-Flash | 93.9% | 2026 |
| NVIDIA | Nemotron 3 Ultra | 86.7% | 2026 |
| MiniMax | MiniMax-M3 | 26.7% | 2026 |
| MiniMax | MiniMax-M3 | 71.1% | 2026 |
| Ai2 (Allen Institute) | Tulu 3 (Tülu 3) 70B | 4.4% | 2024 |
| Microsoft | Phi-4 | 13.8% | 2024 |
Record per lab: FrontierMath (tiers 1 to 3) ?
- Anthropic 95.1%
- OpenAI 93.7%
- Google 86%
- Meta 82.1%
- Moonshot AI 78.2%
- Alibaba 74.7%
- DeepSeek 71.2%
- Zhipu AI (Z.ai) 68.8%
- xAI 66%
- Mistral AI 58.6%
Source: Epoch AI, CC BY 4.0, latest version we checked. The 90% to 100% band marks the saturation zone.
Chart data table
| Lab | Model | % | Dates in sources |
|---|---|---|---|
| Anthropic | Claude Opus 4.1 | 12.6% | 2025 |
| Anthropic | Claude Sonnet 4.5 | 23.9% | 2025 |
| Anthropic | Claude Opus 4.5 | 34.4% | 2025 |
| Anthropic | Claude Opus 4.6 | 66% | 2026 |
| Anthropic | Claude Opus 4.7 | 70.2% | 2026 |
| Anthropic | Claude Opus 4.8 | 80% | 2026 |
| Anthropic | Claude Fable 5 | 87% | 2026 |
| Anthropic | Claude Fable 5.1 | 90.2% | 2026 |
| Anthropic | Claude Opus 5.5 | 91.2% | 2026 |
| Anthropic | Claude Fable 5.2 | 95.1% | 2026 |
| Moonshot AI | Kimi K2.6 | 57.2% | 2026 |
| Moonshot AI | Kimi K3 | 72.2% | 2026 |
| Moonshot AI | Kimi K3.1 Pro | 78.2% | 2026 |
| OpenAI | GPT-3.5 Turbo | 0% | 2024 |
| OpenAI | GPT-4 Turbo (Apr 2024) | 0.7% | 2024 |
| OpenAI | o1 | 8.4% | 2024 |
| OpenAI | o1 | 10.2% | 2024 |
| OpenAI | o1 | 14.7% | 2024 |
| OpenAI | o3-mini | 18.6% | 2025 |
| OpenAI | o3 | 19.3% | 2025 |
| OpenAI | o4-mini | 28.8% | 2025 |
| OpenAI | o3 | 29.8% | 2025 |
| OpenAI | o3 | 33.3% | 2025 |
| OpenAI | o4-mini | 36.1% | 2025 |
| OpenAI | GPT-5 | 37.2% | 2025 |
| OpenAI | GPT-5 mini | 46.7% | 2025 |
| OpenAI | GPT-5 | 55.4% | 2025 |
| OpenAI | GPT-5 Pro | 55.8% | 2025 |
| OpenAI | GPT-5.2 | 67.4% | 2025 |
| OpenAI | GPT-5.2 Pro | 74% | 2025 |
| OpenAI | GPT-5.4 | 78.6% | 2026 |
| OpenAI | GPT-5.4 Pro | 82.5% | 2026 |
| OpenAI | GPT-5.5 | 85.3% | 2026 |
| OpenAI | GPT-5.5 Pro | 87.7% | 2026 |
| OpenAI | GPT-5.6 Sol | 89.1% | 2026 |
| OpenAI | GPT-6 Astra | 93.7% | 2026 |
| DeepSeek | DeepSeek-V4-Pro | 45.3% | 2026 |
| DeepSeek | DeepSeek V4 Flash 0731 | 57.5% | 2026 |
| DeepSeek | DeepSeek V4 Pro 0813 | 64.6% | 2026 |
| DeepSeek | DeepSeek V4.1 Flash | 67.4% | 2026 |
| DeepSeek | DeepSeek V4.1 Flash 1017 | 71.2% | 2026 |
| Meta | Muse Spark 1.3 | 74% | 2026 |
| Meta | Muse Spark 1.3 | 74.4% | 2026 |
| Meta | Muse Wonder | 82.1% | 2026 |
| Gemini 2.5 Pro (Jun 2025) | 24.6% | 2025 | |
| Gemini 3 Flash | 51.2% | 2025 | |
| Gemini 3.1 Pro | 59.6% | 2026 | |
| Gemini 3.5 Flash | 62.8% | 2026 | |
| Gemini 3.7 Flash | 71.6% | 2026 | |
| Gemini 4 Argon | 86% | 2026 | |
| Mistral AI | Mistral Large 4 | 58.6% | 2026 |
| xAI | Grok 4.20 | 44.9% | 2026 |
| xAI | Grok 4.5 | 57.2% | 2026 |
| xAI | Grok 4.6 | 66% | 2026 |
| Alibaba | Qwen3-Max | 18.9% | 2025 |
| Alibaba | Qwen3.5 397B-A17B | 29.5% | 2026 |
| Alibaba | Qwen3.5 397B-A17B | 31.2% | 2026 |
| Alibaba | Qwen 3.6 Plus | 32.3% | 2026 |
| Alibaba | Qwen 3.6 Plus | 38.2% | 2026 |
| Alibaba | Qwen3.7-Max | 64.6% | 2026 |
| Alibaba | Qwen 3.8 Max | 74.7% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.1 | 24.9% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.1 | 36.8% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 42.5% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 54.7% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 59.2% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.3 | 68.8% | 2026 |
Record per lab: SimpleQA Verified ?
- Anthropic 94%
- Google 76%
- OpenAI 75.6%
- Moonshot AI 62.2%
- Meta 60.3%
- xAI 56%
- Alibaba 55.8%
- DeepSeek 52.9%
- ByteDance 44.7%
- Zhipu AI (Z.ai) 41%
- Mistral AI 22.4%
Source: Epoch AI, CC BY 4.0, latest version we checked. The 90% to 100% band marks the saturation zone.
Chart data table
| Lab | Model | % | Dates in sources |
|---|---|---|---|
| Anthropic | Claude 3 Opus | 12.6% | 2024 |
| Anthropic | Claude Sonnet 4.5 | 23.7% | 2025 |
| Anthropic | Claude Sonnet 4.5 | 30.7% | 2025 |
| Anthropic | Claude Opus 4.5 | 45.7% | 2025 |
| Anthropic | Claude Opus 4.6 | 47% | 2026 |
| Anthropic | Claude Opus 4.7 | 51.7% | 2026 |
| Anthropic | Claude Opus 4.8 | 53% | 2026 |
| Anthropic | Claude Fable 5 | 70.7% | 2026 |
| Anthropic | Claude Fable 5.1 | 70.8% | 2026 |
| Anthropic | Claude Opus 5.5 | 72.2% | 2026 |
| Anthropic | Claude Fable 5.2 | 94% | 2026 |
| Moonshot AI | Kimi K2.5 | 34.3% | 2026 |
| Moonshot AI | Kimi K2.6 | 34.9% | 2026 |
| Moonshot AI | Kimi K2.7 Code | 36.5% | 2026 |
| Moonshot AI | Kimi K3 | 50.6% | 2026 |
| Moonshot AI | Kimi K3.1 Pro | 62.2% | 2026 |
| ByteDance | Doubao-Seed-1.6 | 17.4% | 2025 |
| ByteDance | Doubao-Seed-2.0 Pro | 39.4% | 2026 |
| ByteDance | Doubao Seed 2.1 Pro | 44.7% | 2026 |
| Mistral AI | Mistral Large 4 | 20% | 2026 |
| Mistral AI | Mistral Large 4 | 22.4% | 2026 |
| OpenAI | GPT-4o mini | 8.3% | 2024 |
| OpenAI | GPT-4o | 26% | 2024 |
| OpenAI | o1 | 41.1% | 2024 |
| OpenAI | o3 | 49.4% | 2025 |
| OpenAI | GPT-5 | 50.1% | 2025 |
| OpenAI | GPT-5.5 | 63% | 2026 |
| OpenAI | GPT-5.6 Sol | 69.7% | 2026 |
| OpenAI | GPT-6 Astra | 75.6% | 2026 |
| DeepSeek | DeepSeek-V4-Pro | 47% | 2026 |
| DeepSeek | DeepSeek V4 Pro 0813 | 52.9% | 2026 |
| Meta | Muse Spark 1.1 | 57.8% | 2026 |
| Meta | Muse Spark 1.2 | 60.3% | 2026 |
| Gemini 3 Flash | 66.8% | 2025 | |
| Gemini 3.1 Pro | 73.5% | 2026 | |
| Gemini 4 Argon | 76% | 2026 | |
| xAI | Grok 4.20 | 30.2% | 2026 |
| xAI | Grok 4.3 Beta | 33.2% | 2026 |
| xAI | Grok 4.5 | 48.3% | 2026 |
| xAI | Grok 4.6 | 48.9% | 2026 |
| xAI | Grok 4.6 | 49.3% | 2026 |
| xAI | Grok 4.7 | 56% | 2026 |
| Alibaba | Qwen3-235B-A22B (Jul 2025) | 40.4% | 2025 |
| Alibaba | Qwen3-Max | 48.7% | 2025 |
| Alibaba | Qwen 3.6 Max (Preview) | 52% | 2026 |
| Alibaba | Qwen3.7-Max | 55.8% | 2026 |
| Zhipu AI (Z.ai) | GLM-4.7 | 32.2% | 2025 |
| Zhipu AI (Z.ai) | GLM-5.1 | 34% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 34.2% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.3 | 41% | 2026 |
Record per lab: SWE-bench Verified ?
- Anthropic 83.5%
- OpenAI 80.6%
- Google 79.3%
- Zhipu AI (Z.ai) 78.7%
- DeepSeek 77.6%
- Alibaba 77.3%
- Moonshot AI 76.7%
Source: Epoch AI, CC BY 4.0, latest version we checked. The 90% to 100% band marks the saturation zone.
Chart data table
| Lab | Model | % | Dates in sources |
|---|---|---|---|
| Zhipu AI (Z.ai) | GLM-5 | 72.1% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.1 | 74.2% | 2026 |
| Zhipu AI (Z.ai) | GLM-5.2 | 78.7% | 2026 |
| Alibaba | Qwen 3.6 Plus | 57.9% | 2026 |
| Alibaba | Qwen 3.6 Max (Preview) | 76.7% | 2026 |
| Alibaba | Qwen3.7-Max | 77.3% | 2026 |
| DeepSeek | DeepSeek-V4-Pro | 77.6% | 2026 |
| Gemini 2.5 Pro (Jun 2025) | 57.6% | 2025 | |
| Gemini 3 Pro | 72.9% | 2025 | |
| Gemini 3 Flash | 75.4% | 2025 | |
| Gemini 3.1 Pro | 75.6% | 2026 | |
| Gemini 3.5 Flash | 79.3% | 2026 | |
| Moonshot AI | Kimi K2.5 | 73.8% | 2026 |
| Moonshot AI | Kimi K2.6 | 76.7% | 2026 |
| OpenAI | GPT-4o | 31% | 2024 |
| OpenAI | GPT-4.1 | 48.5% | 2025 |
| OpenAI | o3 | 62.3% | 2025 |
| OpenAI | GPT-5 mini | 64.7% | 2025 |
| OpenAI | GPT-5 | 71.5% | 2025 |
| OpenAI | GPT-5 | 73.6% | 2025 |
| OpenAI | GPT-5.2 | 73.8% | 2025 |
| OpenAI | GPT-5.3 Codex | 74.8% | 2026 |
| OpenAI | GPT-5.4 | 76.9% | 2026 |
| OpenAI | GPT-5.5 | 80.6% | 2026 |
| Anthropic | Claude 3.7 Sonnet | 61% | 2025 |
| Anthropic | Claude Opus 4 | 70.7% | 2025 |
| Anthropic | Claude Opus 4.1 | 73.3% | 2025 |
| Anthropic | Claude Opus 4.5 | 76.7% | 2025 |
| Anthropic | Claude Opus 4.6 | 78.7% | 2026 |
| Anthropic | Claude Opus 4.7 | 83.5% | 2026 |
Input price per 1M tokens by OpenRouter listing date (log scale) ?
- Alibaba
- Anthropic
- DeepSeek
- Meta
- Mistral AI
- Moonshot AI
- OpenAI
- Zhipu AI (Z.ai)
- xAI
Source: OpenRouter, latest version we checked. Dates come from the sources: OpenRouter listing or Hugging Face repo creation, which may differ from the official launch.
Sources
- Epoch AI, CC BY 4.0, latest version we checked.
- OpenRouter, dados públicos da API; exports de benchmarks e rankings em CC BY 4.0, latest version we checked.
Dates come from the sources: OpenRouter listing or Hugging Face repo creation, which may differ from the official launch. Family and version were grouped by rule from the name.