MCP-Atlas
17 models tested · Updated 2026-06-09 · Verified sources only
Claude Fable 5 leads at 84.7%
1
Anthropic · Blog/Anthropic · 2026-06-09
Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.
84.7%
2
Moonshot AI · Blog/Moonshot AI · 2026-07-16
500-task public subset, 100-turn limit, Gemini 3.1 Pro judge.
84.2%
3
Google DeepMind · Blog/Google DeepMind · 2026-06-10
Highest among compared models. Up from Gemini 3 Flash (62.0).
83.6%
4
OpenAI · Blog/OpenAI · 2026-07-09
Score as cited in Kimi K3 blog (source: OpenAI official).
83.6%
5
Anthropic · Blog/Anthropic · 2026-06-17
Score as cited in Kimi K3 blog (source: Anthropic official).
83.6%
6
OpenAI · Blog/OpenAI · 2026-06-26
Score as cited in Kimi K3 blog (source: OpenAI official).
82.8%
7
Z.ai · Blog/Z.ai · 2026-06-13
Score as cited in Kimi K3 blog (source: Z.ai official).
82.6%
8
Anthropic · Blog/Google DeepMind · 2026-06-10
Below Gemini 3.5 Flash (83.6) and Gemini 3.1 Pro (78.2).
79.1%
9
Google DeepMind · Blog/Google DeepMind · 2026-06-10
Below Gemini 3.5 Flash (83.6).
78.2%
10
Anthropic · Blog/Z.ai · 2026-06-13
Highest among compared models.
77.8%
11
Z.ai · Blog/Z.ai · 2026-06-13
Public set. Up from GLM-5.1 (71.8). MCP tool-use benchmark.
76.8%
12
Thinking Machines Lab · Blog/ThinkingMachines · 2026-07-15
Beats Inkling slightly on MCP-Atlas.
74.9%
13
MiniMax · Blog/Z.ai · 2026-06-13
Below GLM-5.2 (76.8). Agentic tool use.
74.2%
14
Thinking Machines Lab · Blog/Thinking Machines Lab · 2026-07-20
MCP tool-use; above Nemotron 3 Ultra (44.7) and Kimi K2.5 (64.0). Strong agentic foundation.
74.1%
15
DeepSeek · HuggingFace/deepseek-ai-DeepSeek-V4-Pro · 2026-04-24
Pass@1. MCP-Atlas public eval.
73.6%
16
Anthropic · Blog/Google DeepMind · 2026-06-10
Below Gemini 3.5 Flash (83.6) and Opus 4.7 (79.1).
69.5%
17
Google DeepMind · Blog/Google DeepMind · 2026-06-10
Below Gemini 3.5 Flash (83.6). Big jump for new Flash.
62.0%