← All benchmarks

GAIA2

Agents · 2026-02-12

GAIA2 evaluates assistant agents on multi-turn tasks inside simulated, stateful universes of everyday apps such as email, calendar, messaging, contacts and shopping, where an event daemon injects time-based and environment-driven events mid-episode. The public release has 800 scenarios across five capability splits: execution, search, ambiguity, adaptability and time. An LLM judge compares the agent's tool actions and end state with each scenario's oracle events, and the score is the share of scenarios solved.

Top models (higher is better)