APEX-Agents
Long tasks from investment banking, consulting and corporate law, done with real office tools.
Agent Index weight
8%
in the Knowledge-work agents pillar
Models tested
62
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon Supported Google | 84.2% | source | — | — | |
| 2 | Gemini 4 Argon Supported Google | 82.4% | source | — | — | |
| 3 | Claude Fable 5.2 Supported Anthropic | 80.1% | source | — | — | |
| 4 | Gemini 4 Hydrogen Estimated Google | 78.0% | source | — | — | |
| 5 | Claude Sonnet 5.5 Estimated Anthropic | 75.5% | source | — | — | |
| 6 | Claude Opus 5.5 Supported Anthropic | 73.5% | source | — | — | |
| 7 | Grok 4.7 Estimated xAI | 70.0% | source | — | — | |
| 8 | DeepSeek V4.1 Flash 1017 Estimated DeepSeek | 69.8% | source | — | — | |
| 9 | Claude Fable 5.1 Supported Anthropic | 68.6% | source | — | — | |
| 10 | GPT-6.1 Terra Estimated OpenAI | 68.3% | source | — | — | |
| 11 | Gemini 3.7 Flash Estimated Google | 67.8% | source | — | — | |
| 12 | Muse Wonder Estimated Meta | 66.1% | source | — | — | |
| 13 | Claude Opus 5 Supported Anthropic | 65.8% | source | — | — | |
| 14 | Kimi K3.1 Pro Moonshot AI | 65.5% | source | — | — | |
| 15 | Grok 4.6 Supported xAI | 65.3% | source | — | — | |
| 16 | GPT-6 Astra Supported OpenAI | 64.7% | source | — | — | |
| 17 | Gemini 3.8 Flash Supported Google | 64.3% | source | — | — | |
| 18 | Muse Ember Meta | 64.3% | source | — | — | |
| 19 | Claude Fable 5 Supported Anthropic | 63.6% | source | — | — | |
| 20 | Qwen3.8 Max Preview Estimated Alibaba (Qwen) | 63.3% | source | — | — | |
| 21 | GPT-6.1 Sol Estimated OpenAI | 60.0% | source | — | — | |
| 22 | Muse Spark 1.3 Estimated Meta | 58.6% | source | — | — | |
| 23 | GPT-5.6 Terra Estimated OpenAI | 58.2% | source | — | — | |
| 24 | GLM-5.3 Supported Z.ai (Zhipu) | 56.6% | source | — | — | |
| 25 | Grok 4.5 Estimated xAI | 56.2% | source | — | — | |
| 26 | Claude Haiku 5.5 Estimated Anthropic | 56.1% | source | — | — | |
| 27 | GPT-5.5 Supported OpenAI | 55.1% | source | — | — | |
| 28 | Claude Sonnet 5 Supported Anthropic | 54.5% | source | — | — | |
| 29 | MiMo-V2.6-Pro Xiaomi | 54.4% | source | — | — | |
| 30 | GPT-6 Sol Estimated OpenAI | 54.3% | source | — | — | |
| 31 | GLM-5.3-Flash Supported Z.ai (Zhipu) | 52.8% | source | — | — | |
| 32 | GPT-5.4 Supported OpenAI | 52.4% | source | — | — | |
| 33 | Mistral Large 4 Mistral | 51.6% | source | — | — | |
| 34 | GPT-5.6 Sol Supported OpenAI | 51.4% | source | — | — | |
| 35 | Claude Opus 4.7 Supported Anthropic | 49.2% | source | — | — | |
| 36 | Claude Opus 4.8 Supported Anthropic | 48.9% | source | — | — | |
| 37 | Beam 1.0 Reflection | 48.8% | source | — | — | |
| 38 | Kimi K3 Supported Moonshot AI | 47.8% | source | — | — | |
| 39 | Qwen3.8 27B Alibaba (Qwen) | 47.5% | source | — | — | |
| 40 | DeepSeek V4 Pro 0813 DeepSeek | 47.3% | source | — | — | |
| 41 | Gemini 3.6 Flash Estimated Google | 46.9% | source | — | — | |
| 42 | Inkling 2 Small Estimated thinkingmachines | 46.8% | source | — | — | |
| 43 | Claude Opus 4.6 Supported Anthropic | 46.3% | source | — | — | |
| 44 | Muse Spark 1.2 Estimated Meta | 45.4% | source | — | — | |
| 45 | GLM-5.2 Supported Z.ai (Zhipu) | 45.2% | source | — | — | |
| 46 | GPT-6 Luna OpenAI | 44.3% | source | — | — | |
| 47 | Claude Sonnet 4.6 Supported Anthropic | 43.0% | source | — | — | |
| 48 | GPT-5.6 Luna Estimated OpenAI | 43.0% | source | — | — | |
| 49 | GLM-5.1 Estimated Z.ai (Zhipu) | 40.9% | source | — | — | |
| 50 | DeepSeek V4.1 Flash Estimated DeepSeek | 39.5% | source | — | — | |
| 51 | MiniMax-M3 Estimated MiniMax | 37.7% | source | — | — | |
| 52 | Kimi K2.7 Code Estimated Moonshot AI | 37.6% | source | — | — | |
| 53 | Gemini 3.1 Pro Preview Supported Google | 35.1% | source | — | — | |
| 54 | Inkling Estimated thinkingmachines | 33.8% | source | — | — | |
| 55 | Muse Spark Open Meta | 33.4% | source | — | — | |
| 56 | Muse Spark 1.1 Estimated Meta | 31.8% | source | — | — | |
| 57 | Gemini 3.5 Flash Lite Google | 29.3% | source | — | — | |
| 58 | Gemini 3.5 Flash Supported Google | 27.5% | source | — | — | |
| 59 | Qwen3.5 397B-A17B Alibaba (Qwen) | 24.9% | source | — | — | |
| 60 | Nemotron 3 Ultra NVIDIA | 22.7% | source | — | — | |
| 61 | DeepSeek V3.2 Supported DeepSeek | 21.3% | source | — | — | |
| 62 | GPT OSS 120B Supported OpenAI | 4.4% | source | — | — |
For models without a APEX-Agents result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 58 models (residual SD 0.43 on the logit scale). Estimates are labelled and widen the model’s range.