Benchmark
Apex Agents
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.
- Models measured
- 72
- Scores recorded
- 83
- Sources
- 3
- IQ weight
- 2.12
Score distribution
- 50.7%
- Mean
- 51.6%
- Median
- 16.9%
- Std dev
- 27.9%
- 10th percentile
- 70%
- 90th percentile
- 4.4%
- Min
This benchmark has a weight of 2.12 for chat models, measured over 72 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 51 | Inkling 2 Small· Epoch AI Benchmarking Hub data | 46.8% |
| 52 | Claude Opus 4.6max· Epoch AI Benchmarking Hub data | 46.3% |
| 53 | Muse Spark 1.2· Epoch AI Benchmarking Hub data | 45.4% |
| 54 | GLM 5.2· Epoch AI Benchmarking Hub data | 45.2% |
| 55 | Claude Sonnet 5.5medium· Epoch AI Benchmarking Hub data | 44.6% |
| 56 | GPT-6 Lunamax· Epoch AI Benchmarking Hub data | 44.3% |
| 57 | Claude Sonnet 4.6high· Epoch AI Benchmarking Hub data | 43% |
| 58 | GPT-5.6 Luna· Epoch AI Benchmarking Hub data | 43% |
| 59 | Kimi K3max effort · temp 1.0 · top_p 1.0· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 41% |
| 60 | GLM 5.1· Epoch AI Benchmarking Hub data | 40.9% |
| 61 | DeepSeek V4.1 Flash· Epoch AI Benchmarking Hub data | 39.5% |
| 62 | MiniMax M3· Epoch AI Benchmarking Hub dataValues from 2 sources
| 37.7% |
| 63 | Kimi K2.7 Code· Epoch AI Benchmarking Hub data | 37.6% |
| 64 | Hy4 preview· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 37.1% |
| 65 | Hy4 previewHF Eval Result· Reported by Tencent Hy· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 37.1% |
| 66 | Gemini 3.1 Pro Preview· Epoch AI Benchmarking Hub dataValues from 2 sources
| 35.1% |
| 67 | Doubao Seed 2.1 Pro· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 33.8% |
| 68 | Inkling FP4· Epoch AI Benchmarking Hub data | 33.8% |
| 69 | Gemini 3.1 Pro Previewsetting not named by the source· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 33.5% |
| 70 | Muse Spark Open 816B A42B· Epoch AI Benchmarking Hub data | 33.4% |
| 71 | Muse Spark 1.1· Epoch AI Benchmarking Hub data | 31.8% |
| 72 | Gemini 3.5 Flash Lite· Epoch AI Benchmarking Hub data | 29.3% |
| 73 | Doubao Seed 2.1 Turbo· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 29.2% |
| 74 | Ling 3.0 Flash Fin· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 29.2% |
| 75 | Kimi K2.6· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 27.9% |
| 76 | Kimi K2.6Kimi K2.6 official benchmark table · thinking · temp 1.0 · top_p 1.0 · 262,144 context · 452-task subset· Reported by Moonshot AI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 27.9% |
| 77 | MiniMax M3Archipelago · ReAct Toolbelt · Claude Sonnet 4.6 grader· Reported by MiniMax Research· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 27.7% |
| 78 | Gemini 3.5 Flash· Epoch AI Benchmarking Hub data | 27.5% |
| 79 | Hy3 (Deep Infra)· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 25.6% |
| 80 | Qwen3.5 397B A17B· Epoch AI Benchmarking Hub data | 24.9% |
| 81 | Nemotron 3 Ultra· Epoch AI Benchmarking Hub data | 22.7% |
| 82 | DeepSeek V3.2· Epoch AI Benchmarking Hub data | 21.3% |
| 83 | GPT OSS 120B· Epoch AI Benchmarking Hub data | 4.4% |
Showing 51–83 of 83 results.
Data credits: Epoch AI Benchmarking Hub, CC BY 4.0 (licence) · Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.