Devin Fusion: Frontier Performance at 35% Lower Cost

Cognition introduces Devin Fusion, a multi-model agent harness that routes coding tasks between a frontier model ("main agent") and a cheaper "sidekick" model, aiming to preserve frontier-quality results at lower cost.

The key architectural insight is a sidekick approach: both models run in parallel with independent cached contexts. The main agent delegates routine work to the sidekick by default while retaining authority over planning, ambiguity, and final review. This avoids cache misses from model hopping — switching happens at context compaction boundaries, making it effectively free.

On Cognition's FrontierCode benchmark, Fusion without Fable 5 scored 47.9 at $2.38/task versus Opus 4.8's 48.8 at $3.24 and GPT-5.5's 44.8 at $3.64 — a ~35% cost improvement against frontier models. Fusion paired with Fable 5 hit 57.6 at $3.00/task (41% cheaper than pure Fable 5).

Worked examples show cost savings of 25–62% across varied tasks with quality maintained or improved, except when the task demands judgment as the deliverable — delegating then backfires (a React/Redux team-selector task dropped from score 54 to 27).

The post's broader argument: the era of single-model workloads is ending. Rising frontier costs and a growing pool of specialized models make multi-model routing the natural next step, just as you wouldn't use a supercar for a grocery run.

Internally, 88% of merged Cognition PRs were driven entirely by the automated Fusion router.