Claude Fable 5, take two: same model, different harness, and a very different result

Endor Labs, an AI-driven software company, has been testing the capabilities of their frontier model, Fable 5, in various coding tasks. In their first experiment, Fable 5, paired with the Claude Code harness, produced an average scorecard of 59.8% FuncPass and 19.0% SecPass under the fair scoring system. This score did not meet the high expectations of the model's performance, as reported by Anthropic.