Devin launches Fusion mode with two AI models working together
In the new mode, one model plans and reviews the task while another writes code and runs tests.
On September 11, Cognition introduced Fusion mode in Devin’s desktop app and command-line tool. Users can now select two models for a single task: one directs the work and the other carries it out.
How are tasks divided?
The lead model plans the work, clarifies ambiguous requirements and reviews the result. It gives the helper model a task, constraints and success criteria.
The helper examines the code, makes changes, runs tests and reports back. If a problem arises, the lead model can take over again.
Each model keeps its own work history. They exchange tasks, results and feedback instead of copying the entire conversation.
Cost savings depend on the task
In tests published by Cognition, Fusion cost less than some single-model configurations. However, both the savings and output quality varied by test and model pairing.
These results therefore cannot be treated as a guarantee of identical savings in every project. The main idea is to assign complex planning to a capable model and practical execution to a less expensive helper.
Photo
An illustrative photo related to the topic, not an image of the Devin interface. Ilya Pavlov / Unsplash. https://unsplash.com/photos/monitor-showing-java-programming-OqtafYT5kTw
Usage license: https://unsplash.com/license
Sources and image credits
Source
Cognition, September 11, 2026: https://cognition.com/blog/local-fusion
This article was prepared by the Founders Cafe editorial team.
