MCP Atlas refreshed judge (20-turn Opus 4.7 report)
Agents · 2026-04-16
The refreshed-judge and retry-handling MCP-Atlas run reported in the Opus 4.7 system card: 20 turns with 10–25 exposed tools per task. This is distinct from both the original harness and Scale's subsequently documented 100-tool-call configuration. The card does not explicitly identify the evaluated task subset; a publicly visible leaderboard is not evidence of the public dataset split.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.7 | 77.3 |