Cognition launches Devin Fusion, routing coding tasks across models to cut costs 35%

Cognition released Devin Fusion in preview on June 29, 2026, a multi-model harness that routes agentic coding subtasks between a frontier model and a cheaper one. The company says it cuts costs 35% on its FrontierCode Extended benchmark while holding performance steady. The Devin Fusion preview is available through the existing Devin product.

The idea attacks the economics of agentic coding, where running a frontier model on every step gets expensive fast. Devin Fusion uses lightweight classifiers to watch task execution and hand work between a frontier “primary” model and a cheaper “sidekick,” timing the handoff during context compaction to limit cache penalties. The sidekick handles exploration, test writing and lint fixes; the frontier model handles planning and final review.

The cost claims are specific. Against frontier models such as GPT-5.5 and Claude Opus 4.8, Cognition reports Devin Fusion is 35% cheaper at equivalent performance, rising to 41% cheaper when Fable 5 is the frontier component. On FrontierCode Extended, the company reports Fusion paired with Fable 5 scoring 57.6 at $3.00 average cost, versus 47.9 at $2.38 for Fusion alone. The 41% figure applies specifically to the Fable 5 configuration, not across the board.

The numbers come from Cognition’s own testing and have not been independently verified. The company also said 88% of merged pull requests in internal testing were entirely router-driven, an in-house metric on its own benchmark rather than a third-party result.

Devin Fusion is a feature of the existing Devin subscription, and the launch did not disclose new pricing tiers. Whether the routing holds up on real codebases outside Cognition’s benchmark is the test that matters.