Breaking Out of the Lab: Testing AI in Professional Domains

A new benchmark series, PRBench (Professional Reasoning Bench), has been introduced to evaluate the real-world reasoning capabilities of frontier AI models in professional domains like finance and law. Unlike academic-style benchmarks, PRBench uses realistic, open-ended tasks authored by domain experts to assess how models perform in complex, high-stakes scenarios with tangible economic consequences. Initial results show that while top models like GPT-5 Pro perform best, all models struggle with complex professional reasoning, especially on the most difficult tasks. Common failures include making inaccurate judgments, providing opaque reasoning, and lacking domain-specific diligence, indicating that current models are not yet ready for critical professional use.