Wharton professor beats everyone to testing Anthropic's new model: It's almost AGI, and it will build itself a bureaucracy.

Wharton professor Ethan Mollick reported that Anthropic’s unreleased Claude Fable 5.2 model delivered “almost AGI” performance across coding, hardware reverse-engineering, entrepreneurship and humanities tasks, generating a complete short film from a single prompt and writing six playable games for a Codex Micro mini keyboard without public documentation. In a startup test, the model screened 22 business ideas, chose a municipal-procurement intelligence product called Prebid Brief, reviewed 201 U.S. city and county meeting agendas, and projected a 20%–40% success rate against a $20,000 first-year revenue target requiring under 20 minutes of human time per month. The same test exposed “jagged AGI” limits: 28 parallel agents spontaneously produced an 87,500-word rulebook and a three-tier approval system, cutting world output from 211 units per day to 2. Mollick’s verdict is that current models match experts on many computer-based tasks but still lack judgment about when to simplify, and he stopped short of running the startup live over doubts about sustained operational reliability. The takeaway for AI watchers is that raw capability is no longer the binding constraint; orchestration discipline and human oversight are what will determine whether agentic systems translate into dependable commercial output.