Benchmarks › APEX-Agents
APEX-Agents
By Mercor ↗Results by Artificial Analysis ↗
Complete professional services tasks with computer tools
42.4top score · Qwen3.8 Max (0902)
29models ranked
14open-weights models
Pass@1 accuracyhigher is better
Oct 20, 2026last updated
How the test works
1
Task
452 tasks
- What the model doesComplete professional services tasks with computer tools
- Kinds of tasksProfessional services tasks
- In → outText, files → Files
2
Run
3 runs per task
- Agent / harnessStirrup in Archipelago with MCP workplace tools
- Step limit200 steps
- Token budgetNon-reasoning: up to 16,384 output tokens, adjusted to model limits; reasoning: model-specific maximum
3
Score
Pass@1 (solved on the first try) accuracy
- Graded byArchipelago local file grader checks Finish answer and filesystem diff; all rubric items must pass; Gemini 3 Flash (low) judge
- Better isHigher
Sources: artificialanalysis.ai
APEX-Agents leaderboardtop 10 of 29
Full leaderboard with filters →| Model | APEX-Agents | Per task | Price | Speed | Supports | Context |
|---|---|---|---|---|---|---|
42.4thinking | $5.41 | $2 / $6 | 40 | 984K | ||
41.3max | $2.00 | $3 / $15 | 112 | 1M | ||
38.9max | $1.40 | $2 / $12 | 142 | 1.05M | ||
38thinking | $1.03 | $1 / $2.70 | 93 | 1M | ||
37.7xhigh | $2.63 | $5 / $30 | 113 | 1.05M | ||
35.8max | $0.18 | $0.20 / $1.20 | 178 | 1.05M | ||
33.7max | $1.47 | $1.40 / $4.40 | 189 | 1M | ||
33.3xhigh | — | $2.50 / $15 | 127 | 1.05M | ||
33max | — | $5 / $25 | 48 | 1M | ||
32thinking | $1.30 | $2 / $12 | 115 | 1M |
APEX-Agents questions
What is APEX-Agents?›
Complete professional services tasks with computer tools. It has 452 tasks. Created by Mercor.
Which model is best on APEX-Agents?›
Qwen3.8 Max (0902) leads APEX-Agents with 42.4, ahead of Kimi K3 (41.3) and GPT-5.6 Terra (38.9).
What is the best open-weights model on APEX-Agents?›
Kimi K3 with 41.3, then Step 5 Preview (38).
How is APEX-Agents scored?›
Pass@1 (solved on the first try) accuracy; archipelago local file grader checks finish answer and filesystem diff; all rubric items must pass; gemini 3 flash (low) judge; 3 runs per task per task.