Fable 5.2 tops our construction benchmarks, but on the main one it was only about 1 percentage point ahead of Opus 5.5. It also costs twice as much. For everyday tasks like reading a drawing or building an estimate, I would stay on Opus.
Fable is built for long jobs. Anthropic pitches it for running complicated work and handing tasks to Sonnet, Haiku and Opus sub-agents. On the Remote Labor Index, which tests long real-world projects, Fable scored 41% and Opus 21.3%.
Fable is more careful about getting the detail right. We had both build a 3D model from 2D warehouse drawings. Opus made it look right. Fable got the data inside the model file right, with concrete within 1% of the true quantity (Opus was 15% out) and the correct number of footings.
My setup for long jobs:
- The AI grills you with questions until you agree what done looks like.
- Fable turns that into steps, each with a test.
- Sonnet or Opus sub-agents do each step and run its test.
- A fresh Fable instance with a short prompt checks the result against the original brief.
I used this to run 5,000 simulations of my construction schedule. It gave us less than a 1% chance of making the finish date, and the critical path was driven more by the external laydown than the building foundations. My original critical path was wrong.
Has anyone found a construction job where Fable is worth the extra cost?