MCP Atlas April 2026 (public 500 tasks)
Agents
Public 500-task subset under Scale's April 2026 revised judge, transient-tool-error retries and 100-tool-call budget. Kept separate from the 1,000-task leaderboard metric. The Inkling-Small source explicitly credits Scale and reports public/all score pairs; these observations retain the public values.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.5 Flash-Lite | 79.8 |
| Inkling-Small | 79.6 |
| Inkling | 78.8 |
| GPT-5.6 Luna | 77.0 |
| Qwen3.5 397B A17B | 74.2 |
| DeepSeek-V4-Flash | 69.0 |
| MiniMax M2.7 | 49.4 |
| Nemotron 3 Ultra 550B A55B | 47.4 |