How to Build Fable, Cheaper Than Opus

When running coding agents, using a more powerful model improves performance but increases the cost burden. To solve this, the Devin Fusion architecture was introduced, using a method where the main model delegates tasks to a cheaper sidekick model.

As a result of experiments, even though Fable 5 costs 2x more per token than Opus 4.8, when used together with a sidekick, the overall cost turned out to be lower. The average cost of the Fable + Sidekick combination is $1.86, which is cheaper than Opus + Sidekick's $2.04, and it also showed higher performance (Score).

This cost reversal phenomenon stems from differences in the models' 'management style.'

• Opus 4.8: Behaves like a micromanager with an intern, directly reading and editing a lot of code and maintaining long context.
• Fable 5: Behaves like a manager with a capable engineer, preemptively blocking unnecessary work and delegating efficiently.

Compared to Opus, Fable 5 has far fewer Lead Turns (number of actions by the main model), and it also drastically reduced input and output token usage. In particular, in 81% of sessions led by Fable, the main model didn't directly edit code at all, demonstrating an efficient workflow.