← All benchmarks

GlobeBench

General QA · 2026-10-15

Patronus AI's benchmark of language models as text world models: given a frozen agent trajectory of actions and environment observations, the model must predict the environment's next observation, which requires tracking earlier state changes, following the environment's rules and propagating side effects. Its 1,632 execution-grounded items span five domains (Code, Enterprise, Games, Math and Science, Dialogue) and 19 simulation skills. A hybrid scorer grades each response: a deterministic grader certifies a pass or a concrete contradiction, and when it abstains GPT-5.6 Terra judges the response at low reasoning effort, with GPT-5.6 Sol as fallback. The score is task success, passes over all 1,632 items with one response per item, in percent; provider content-filter refusals and empty or errored completions count as failures and stay in the denominator.

Top models (higher is better)