Which AI gets real work done?
Agents take many steps on their own: browsing, running code, using apps. Here's who does it best, from real user sessions and from an open, repeatable test suite.
Agent Arena: top 20 models
Scored from 2,711,241 real agent sessions. Score: Agent Arena's rating from user signals, shown as a percentage like on Arena (higher is better). Updated Nov 2, 2026.
Source: Arena (formerly LMArena) leaderboard dataset, Agent Arena board, CC BY 4.0. The chip shows the median cost per task.
Agent setup × model
Exgentic Open Agent Leaderboard: the same 5 models run inside 5 agent setups on AppWorld, BrowseComp+, SWE-bench and Tau² (each counts a quarter). Share of tasks solved.
| Agent setup | DeepSeek V3.2 | Kimi K2.5 | GPT-5.2 | Claude Opus 4.5 | Gemini 3 Pro |
|---|---|---|---|---|---|
| Claude Code CLI | 42% | 30% | 39% | 67% | 56% |
| OpenAI Solo | 32% | 25% | 39% | 73% | 59% |
| SmolAgents Code | 41% | 42% | 38% | 66% | 56% |
| LiteLLM Tool Calling | 46% | 43% | 41% | 61% | 61% |
| LiteLLM Tool Calling with Shortlisting | 45% | 43% | 46% | 62% | 62% |
Best combo: OpenAI Solo with Claude Opus 4.5 (73%). These are older models: the dataset was last updated May 20, 2026. Source: Open Agent Leaderboard results (Exgentic), CDLA-Permissive-2.0.