AI Research Atlas

Devin Fusion

Cognition · 29 June 2026

Devin Fusion routes work between a frontier lead model and a cheaper sidekick, claiming Fable 5-level quality at 60% lower cost.

It uses two strategies. One is a main agent that plans and reviews beside a cheaper executor with separate cached contexts, and the other is classifiers that switch models at context compaction. On FrontierCode it scores 63.1 at $1.35 per task. Fusion reached Devin Desktop and CLI on 2026-09-11.

Date
Monday, 29 June 2026
Lab
Cognition
Kind
feature
Access
research preview

Figures

MeasureValueMeasured by
FrontierCode score / cost63.1 at $1.35 per task
company eval; 25-62% savings by task type
company

On 2026-09-11 Cognition claimed Fusion was up to 39% cheaper than Claude Code or Codex harnesses on its own tests.

Sources

  1. cognition.com/blog/devin-fusion
  2. cognition.com/blog/local-fusion

This record was checked against its sources on 6 October 2026. How we check

Related