Which model leads AgentWorldBench - MCP?

AgentWorldBench - MCP Leaderboard — 14 Models

Pro Mode — raw scores, one benchmark under the lensthird-party eval harness

The field · 14 measured

54.9370.1

14 models measured, most on the third-party eval harness. GPT-5.4 tops the board at 70.1.

14 measured·Lifecycle Updated Oct 8, 2026Score basis

#◆Vendor$/M in$/M out

Leader — ignites whiteField — fill cools with scoreIndependently verified only0 of 14Bars show 54.93–70.1 · ◆ solid = first-party or better · outlined = arm's-length

01GPT-5.4OpenAI70.1$2.5/M$15.0/M$0.125

02Claude Sonnet 4.6Anthropic70$3.0/M$15.0/M$0.129

03Claude Opus 4.6Anthropic69.9$5.0/M$25.0/M$0.215

04Qwen3.5 397B A17BAlibaba68.31$0.60/M$3.6/M$0.031

05Qwen AgentWorld 397B A17BAlibaba68.24———

06GLM-5.1Z.ai67.6$1.4/M$4.4/M$0.043

07Kimi K2.6Moonshot65.23$0.95/M$4.0/M$0.038

08Qwen AgentWorld 35B A3BAlibaba64.79———

09DeepSeek-V4-ProDeepSeek63.27$0.43/M$0.87/M$0.010

10Gemini 3.1 ProGoogle59.07$2.0/M$12.0/M$0.119

11Qwen3.5 35B A3BAlibaba57.87$0.25/M$2.0/M$0.019

12MiniMax M2.7MiniMax55.82$0.21/M$0.84/M$0.009

13Qwen3.6 PlusAlibaba55.28$0.50/M$3.0/M$0.032

14Claude Opus 4.8Anthropic54.93$5.0/M$25.0/M$0.273

◆ solid = independently verified · aggregator attested · vendor attributed — outlined = cross-reference · source-attributed. Dimmed ◆ = confidence inferred. “+N alts” = alternate disclosures in the drawer with why-demoted. Prices per million tokens; $/point = input price ÷ score. Every row opens its model dossier.

Verification: 14 models scored · 0 independently verified · 9 vendor cross-reference · 5 vendor-reported · 0 with source disagreement. How these tiers are assigned