APEX-Agents-AA (avg@3)
Professional Work · 2026-01-20
Artificial Analysis' APEX-Agents-AA protocol evaluates 452 long-horizon professional-services tasks with three independent attempts per task (1,356 scored attempts) using the Stirrup/Archipelago agent harness. The reported pass@1 is the repeated-run mean.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash | 47.1 |
| GPT-5.6 Terra | 38.9 |
| GPT-5.5 | 37.7 |
| GPT-5.6 Luna | 35.8 |
| GLM-5.2 | 33.7 |
| GPT-5.4 | 33.3 |
| Opus 4.6 | 33.0 |
| Gemini 3.1 Pro Preview | 32.0 |
| Kimi K2.6 | 28.5 |
| GPT-5.4 Mini | 28.2 |
| Sonnet 4.6 | 28.0 |
| Gemini 3 Flash Preview | 27.7 |
| GPT-5.4 Nano | 24.9 |
| DeepSeek-V4-Pro | 24.3 |
| Qwen3.7-Plus | 22.4 |
| Grok 4.3 | 17.0 |
| Qwen3.5 397B A17B | 15.3 |
| Step 3.7 Flash | 14.8 |
| DeepSeek-V3.2 | 14.5 |
| GLM-5 | 14.5 |
| Grok 4.20 | 14.2 |
| Gemini 3.1 Flash-Lite Preview | 12.2 |
| Kimi K2.5 | 11.5 |
| MiniMax M2.7 | 10.6 |
| gpt-oss-120b | 3.1 |
| MiMo-V2.5-Pro | 2.4 |
| Nemotron 3 Super 120B A12B | 1.8 |
| gpt-oss-20b | 0.7 |