MCP Atlas April 2026 (all 1,000 tasks)
Agents
Scale's revised MCP-Atlas pass rate over all 1,000 public and private tasks. The April update changes the judge, retries transient tool errors, and replaces the former 20-turn cap with at most 100 tool calls per task. Passing requires mean claim coverage of at least 0.75. All current Scale leaderboard models were re-scored with this protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark Open | 89.2 |
| Muse Spark 1.1 | 88.1 |
| Claude Fable 5.1 | 87.2 |
| Claude Opus 5 | 85.8 |
| GLM-5.3 | 84.2 |
| Gemini 3.5 Flash | 83.6 |
| Claude Fable 5 | 83.3 |
| Kimi K3 | 82.3 |
| Opus 4.8 | 82.2 |
| Muse Spark | 82.2 |
| GPT-5.6 Sol | 81.8 |
| Inkling-Small | 79.2 |
| Opus 4.7 | 79.1 |
| Gemini 3.1 Pro Preview | 78.2 |
| GLM-5.2 | 77.8 |
| Opus 4.6 | 76.8 |
| Inkling | 76.0 |
| GLM-5.1 | 75.6 |
| GPT-5.5 | 75.3 |
| GPT-5.4 | 70.6 |
| Gemini 3 Pro Preview | 70.3 |
| Opus 4.5 | 69.8 |
| Sonnet 4.6 | 69.5 |
| GPT-5.2 | 67.6 |
| Kimi K2.5 | 64.4 |
| Gemini 3 Flash Preview | 62.0 |
| Sonnet 4.5 | 59.5 |
| GLM-4.7 | 58.1 |
| Gemini 3.1 Flash-Lite Preview | 57.1 |
| GPT-5.4 Mini | 56.7 |
| GPT-5.1 | 50.1 |
| o3-pro | 44.5 |
| Haiku 4.5 | 40.2 |