Drew Breunig: Fable Ended the Free Lunch, So Harness and Context Engineering Now Pay Off the Way Optimization Did After Moore's Law Stalled
Breunig argues that before Fable it was rational to skip harness and context work because a cheaper, better model would arrive in months and paper over your problems, exactly like the Herb Sutter "free lunch" of single-threaded CPU scaling. Fable broke that: it is excellent but expensive enough that Opus, GPT-5.6, K3 and GLM are good enough for most code, so teams now have to decide what work goes where. He points at GLM 5.2, released the same week as Fable at roughly 1/9th the cost and about 1/5th the cost of Opus 5, and describes his own split of using Fable to interrogate and shape a design and then handing a brief to GLM. He also expects Fable's access controls, dynamic degradation and required data retention to lock in the multi-model split by pushing companies to think about where their traces go.
↳ Follow the thread