← @bgSponge

Introducing Fusion in Devin CLI

Summary

Cognition introduces Fusion in Devin CLI, a harness for Fable and Astra that reduces coding benchmark costs by up to 39% while maintaining frontier performance.

  1. Fusion splits planning and execution.

    Users select a favorite model to act as the lead for planning and a cheaper model to serve as the sidekick for execution tasks.

    Pick your favorite model for planning and a cost-effective model for execution.

  2. The lead model maintains control.

    Unlike simple routing, the lead reviews all work and can intervene if the sidekick fails to handle a task.

    The lead reviews work, flags problems, and can take control back

  3. Fusion reduces benchmark costs.

    Across coding benchmarks, the Fusion harness achieves savings of 36% to 39% compared to using frontier models alone.

    39% cheaper across coding benchmarks

  4. Performance remains near frontier levels.

    In the Artificial Analysis Coding Agent Index, Devin Fusion Fable 5.1 scores 61.7, nearly matching the 62.2 score of Fable 5.1.

Highlights

  • The most efficient frontier harness for Fable & Astra; 39% cheaper across coding benchmarks.

  • Unlike plain model routing, Fusion keeps frontier intelligence in charge. We don’t rely on one initial routing decision.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →