← All benchmarks

Cresta-Bench

Agents

Cresta-Bench measures how reliably a model runs a contact-center AI agent: four synthetic agents for real-estate lending, a senior-health callback line, a health-plan provider line, and airline support take calls from an LLM-simulated caller across 181 private test cases. Cresta runs each case three times per model and grades runs with deterministic tool-call checks plus LLM judges. The score is pass^3, the share of cases whose three runs all pass, computed per agent and averaged across the four agents.

Top models (higher is better)