Enterprise AI in Numbers · Focus area
AI agents in numbers
How capable and how connected are AI agents, and how much do they already act for people?
The longest tasks agents complete half the time now take people 17.4 hours, by METR's measure, and Gemini 4 Argon passes 82.4% of APEX-Agents professional tasks. Real jobs are harder: the best agent delivers only 21.3% of paid freelance projects acceptably and completes 31.4% of OSWorld 2.0 computer tasks. MCP SDK downloads reached 462,039,621 in September 2026.
7 indicators · Updated
Which model leads which job
The best independent result on each benchmark so far, and the model and lab that hold it.
| Job | Leading model | Score | Since | Evaluator |
|---|---|---|---|---|
| Professional work leader | Gemini 4 Argon Google DeepMind | 82.4% | September 30, 2026 | Epoch AI, Benchmarking hub |
| Remote work leader | Claude Opus 5.5 Anthropic | 21.3% | September 22, 2026 | Epoch AI, Benchmarking hub |
| Computer use leader | Claude Opus 5 Anthropic | 31.4% | July 24, 2026 | Epoch AI, Benchmarking hub |
More indicators
Agent pull requests
1,100,065
Act now
MCP SDK downloads
462,039,621
Evaluate
AI agents acting for users
3.9%
Monitor
AI agent task length
17.4 hours
claude-mythos-preview-early
Act now