Beats Claude 4.0 Sonnet and GPT-5 on Unmodified Tau-bench, but can also take massive actions with low hallucination on data.