Agent Arena💼Work

Dynamic ranking of models on how well they orchestrate tools for real-world agentic tasks, based on signals like tool reliability, task completion, and steerability.

Oct 2, 2026

1,294,125 sessions

51 models

Pareto Frontier

Net Improvement x Cost/task from real Agent Mode tasks

CtrlScroll

Pareto Optimal Models

Net Improvement x Cost/task

Claude Fable 5.1 (Max)

Anthropic · Proprietary

+15.92%

$9.08/task

Claude Opus 5.5 (High)

Anthropic · Proprietary

+13.92%

$2.50/task

GPT 6.1 Sol (Max)

OpenAI · Proprietary

+12.18%

$0.68/task

Deepseek V4.1 Flash (Max)

DeepSeek · MIT

+4.05%

$0.12/task

GPT 6 Luna (Max)

OpenAI · Proprietary

+1.90%

$0.06/task

MiMo V2.6 Flash

Xiaomi · MIT

-0.37%

$0.04/task

Cost per Task

Median cost per task (last 14 days)

1Claude Fable 5.1 (Max)$9.08/task+15.92%

2Claude Opus 5.5 (High)$2.50/task+13.92%

3GPT 6.1 Sol (Max)$0.68/task+12.18%

4GPT 6 Astra (Max)$5.86/task+11.91%

5Claude Sonnet 5.5 (Max)$4.36/task+11.65%

6GPT 6 Sol (Max)$1.32/task+9.09%

7Claude Opus 5 (High)$4.61/task+8.75%

8Claude Fable 5 (High)$3.95/task+8.64%

9GPT 5.6 Sol (xHigh)$1.72/task+7.38%

10Claude Opus 5 (Max)$7.77/task+7.03%

11Claude Opus 4.8 (High)$1.46/task+6.92%

12Gemini 4 Argon (High)$0.81/task+6.67%

13Claude Sonnet 5 (High)$1.28/task+5.68%

14GPT 5.5 (xHigh)$1.51/task+4.98%

15GLM 5.2 (Max)$1.15/task+4.45%