Post by Brendan Foody

AI agents have entered the workforce. But can they actually do the job? Introducing APEX-Agents, our new benchmark that evaluates whether agents can complete real, long-horizon professional services deliverables in Google Workspace. We surveyed hundreds of professionals from McKinsey, Goldman Sachs, and Cravath to understand how they spend their time. Then, we worked with them to create 33 scenarios and 480 tasks across investment banking, consulting, and corporate law. This work reflects the deliverables that a professional would spend hours or days completing. We partnered with Box to simulate realistic enterprise file systems by mapping rigorous, domain-specific challenges to the complex structures professionals use every day. We also worked with Harvey to validate that the legal tasks reflect the work of top lawyers at Fortune 500 enterprises and law firms. Then we tested today’s best models. Mercor APEX-Agents Pass@1 leaderboard:  🥇 Google DeepMind Gemini 3 Flash (High): 24.0% 🥈 OpenAI GPT-5.2 (High): 23.0% 🥉 Anthropic Claude Opus 4.5 (High): 18.4%

Video Content