Leaderboard
Open-weight models ranked on GMP knowledge and task completion. Self-hostable models have 70 billion parameters or fewer, so a single GPU server can run them. Cloud-scale models are still open weights, but they need datacentre hardware. Closed-weight models are listed once they have been scored on the current test suite, so the ranking compares like with like; use the Weights filter to show them on their own. Click a model name to view detailed results.
Task type
Creator
Weights
Deployment
| # | Model | Deployment | Overall | GMP Knowledge | Task Completion | Multimodal | Avg Latency | Cost / task | Total Tokens | # Evals |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | Closed | 90.4% | 100.0% | 88.0% | 85.1% | 89.0s | $0.578 | 496k | 65 |
| 2 | GLM-5.3-Flash | Cloud-scale | 90.4% | 100.0% | 88.0% | 79.0% | 168.4s | $0.0052 | 487k | 65 |
| 3 | Hy4 Preview | Cloud-scale | 90.0% | 100.0% | 87.5% | n/a | 128.1s | $0.054 | 644k | 62 |
| 4 | Qwen3.8 Max (2.4T-A95B) | Cloud-scale | 89.4% | 93.2% | 88.5% | n/a | 123.6s | $0.158 | 531k | 62 |
| 5 | Claude Sonnet 5 | Closed | 89.1% | 100.0% | 86.4% | 76.8% | 31.1s | $0.101 | 243k | 65 |
| 6 | GPT-5.6 Sol | Closed | 89.0% | 100.0% | 86.3% | 78.1% | 29.3s | $0.146 | 176k | 65 |
| 7 | Kimi K3 | Cloud-scale | 89.0% | 97.7% | 86.8% | 80.4% | 153.1s | $0.230 | 366k | 65 |
| 8 | GPT-5.6 Terra | Closed | 88.6% | 100.0% | 85.8% | 76.9% | 27.2s | $0.075 | 155k | 65 |
| 9 | GLM-5.3 | Cloud-scale | 88.6% | 95.5% | 86.9% | n/a | 37.0s | $0.050 | 243k | 61 |
| 10 | Qwen3.8-Flash | Cloud-scale | 88.4% | 93.2% | 87.2% | 86.6% | 87.6s | $0.012 | 611k | 65 |
| 11 | Qwen3.8 27B | Self-hostable | 87.6% | 95.5% | 85.7% | 78.8% | 124.5s | $0.062 | 551k | 65 |
| 12 | GPT-5.6 Luna | Closed | 87.5% | 100.0% | 84.3% | 75.9% | 17.8s | $0.0067 | 145k | 65 |
| 13 | MiniMax M3 | Cloud-scale | 86.2% | 95.5% | 83.8% | 62.4% | 29.3s | $0.013 | 270k | 65 |
| 14 | GLM-5.2 | Cloud-scale | 84.0% | 97.7% | 80.5% | n/a | 70.3s | $0.025 | 234k | 62 |
| 15 | Qwen3.6 27B | Self-hostable | 83.9% | 90.9% | 82.1% | 76.1% | 95.0s | $0.020 | 285k | 65 |
| 16 | DeepSeek V4 Pro | Cloud-scale | 82.2% | 93.2% | 79.5% | n/a | 45.4s | $0.018 | 168k | 62 |
| 17 | Qwen3.5-397B-A17B | Cloud-scale | 82.1% | 100.0% | 77.6% | 72.3% | 57.8s | $0.021 | 256k | 65 |
| 18 | Qwen3.6 35B A3B | Self-hostable | 80.7% | 90.9% | 78.2% | 66.5% | 36.8s | $0.0082 | 316k | 65 |
| 19 | DeepSeek V4 Flash | Cloud-scale | 80.4% | 93.2% | 77.3% | n/a | 33.4s | $0.0019 | 152k | 62 |
| 20 | Muse Glimmer 30B | Self-hostable | 80.3% | 84.1% | 79.3% | 74.1% | 33.0s | $0.0066 | 190k | 65 |
| 21 | Qwen3.5-35B-A3B | Self-hostable | 77.9% | 90.9% | 74.7% | 70.8% | 36.6s | $0.0075 | 294k | 65 |
| 22 | MiniMax M2.7 | Cloud-scale | 77.5% | 97.7% | 72.5% | n/a | 42.5s | $0.011 | 236k | 62 |
| 23 | Claude Haiku 4.5 | Closed | 75.7% | 95.5% | 70.7% | 61.4% | 23.0s | $0.050 | 230k | 65 |
| 24 | DeepSeek-V3.2 | Cloud-scale | 75.5% | 93.2% | 71.1% | n/a | 26.7s | $0.0010 | 70k | 62 |
| 25 | DeepSeek-R1 | Cloud-scale | 74.9% | 95.5% | 69.8% | n/a | 108.1s | $0.012 | 163k | 62 |
| 26 | Gemma 4 31B IT | Self-hostable | 74.3% | 97.7% | 68.4% | 70.4% | 24.2s | $0.0010 | 76k | 65 |
| 27 | Gemma 4 26B A4B IT | Self-hostable | 73.6% | 95.5% | 68.2% | 59.2% | 12.5s | <$0.001 | 79k | 65 |
| 28 | Mistral Large 3 675B | Cloud-scale | 72.6% | 97.7% | 66.3% | 45.1% | 13.6s | $0.0047 | 102k | 65 |
| 29 | Mistral Small 2603 | Self-hostable | 70.5% | 90.9% | 65.3% | 29.7% | 7.7s | $0.0018 | 101k | 65 |
| 30 | Llama 4 Maverick | Cloud-scale | 59.3% | 93.2% | 50.8% | 32.2% | 20.9s | $0.0011 | 72k | 65 |
| 31 | Llama 3.3 70B Instruct | Self-hostable | 58.9% | 93.2% | 50.4% | n/a | 20.8s | <$0.001 | 56k | 62 |
| 32 | Llama 4 Scout | Cloud-scale | 56.2% | 93.2% | 47.0% | 26.4% | 16.2s | <$0.001 | 66k | 65 |