MCP Atlas (original protocol)
Agents · 2025-09-19
Historical MCP-Atlas results from the original evaluation, before the April 2026 judge and retry revision. Includes the original Opus 4.5 62.3 and Opus 4.6 59.5/62.7 results and contemporaneous release-table comparisons. The exact evaluated subset is not consistently disclosed, so this cohort does not assert the 500-task public denominator. Revised harnesses and explicitly different source protocols have separate benchmark identities.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.4 | 67.2 |
| Opus 4.6 | 62.7 |
| Opus 4.5 | 62.3 |
| Sonnet 4.6 | 61.3 |
| GPT-5.2 | 60.6 |
| GPT-5.4 Mini | 57.7 |
| Gemini 3 Flash Preview | 57.4 |
| GPT-5.4 Nano | 56.1 |
| Gemini 3 Pro Preview | 54.1 |
| GPT-5 Mini | 47.6 |
| Sonnet 4.5 | 43.8 |
| Gemini 2.5 Pro | 8.8 |
| Gemini 2.5 Flash | 3.4 |