World Agent: leaderboard
Tests whether a language-model agent can keep a delivered interactive world running: over rounds it uses world-interface tools to realize the contracted events, structures and probes of a world specification. Scored by the maintenance completion score, the share of contract items realized, higher is better (World Agent, 2026).
Metric: Maintenance completion score (%): share of a world's contracted events, structures and probes realized, mean per run over the nine-case common subset (27 runs per model, easy/hard/extreme tiers, PhyAgentOS harness, isolated LLM judge plus programmatic audit); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max (0902) | 81.3 | #50 |
| 2 | GLM-5.3 Flash | 75.2 | #122 |
| 3 | DeepSeek V4.1 Flash | 74.4 | #59 |
| 4 | Qwen 3.8 Flash | 73.5 | #112 |
| 5 | GPT-5.6 Luna | 73.4 | #121 |
| 6 | Kimi K3 | 70.9 | #54 |
| 7 | Gemini 3.8 Flash | 70.4 | #36 |
| 8 | GLM-5.3 | 69.2 | #66 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=world-agent · How It Works · Data refreshed daily, snapshot 2026-11-13.