Fusion Lands in Devin Desktop and CLI: Cognition's Multi-Model Coding Harness
Cognition has brought its multi-model coding harness Fusion to Devin Desktop and CLI. A Lead model handles planning and review while a Sidekick model handles ex
On September 12, 2026, Cognition introduced 'Fusion', a multi-model coding harness, to the Devin Desktop and CLI environments. Instead of assigning every step to a single model, users pair a Lead model for planning and code review with a Sidekick model for execution. According to the company, this structure maintains frontier-level benchmark performance while lowering task costs.
Image source: https://cognition.com/blog/local-fusion
This release targets Devin Desktop and CLI users. Asked about web app support, Nader Dabit replied "Unsure at the moment" in a social post thread, so expansion to the web app is not yet confirmed.
What Fusion Does: Splitting Work Between Lead and Sidekick
With Fusion, you pick two models instead of one. A frontier model takes the 'Lead' role for planning and review, and a cost-effective model takes the 'Sidekick' role for execution. For best results, Cognition recommends pairing Lead 'Fable 5.1' with Sidekick 'SWE-2'.
- Lead (planning and review) options: Fable, Astra, Sol, Opus — Cognition says it evaluated Fable 5.1 and GPT-6 Astra each paired with SWE-2.
- Sidekick (execution) options: SWE-2 Free, Luna, Sol, GLM Free
- Speed: Normal or Fast
- Reasoning level: adjustable per task
It is most useful for Devin Desktop and CLI users who want to keep frontier-model planning and review quality while spending less on execution. How to balance delegation and token spend between the higher-cost Lead and the free or low-cost Sidekick was raised as a key question in a community reply — reported here as a third-party observation, not as an official Cognition precondition.
Reported Results: The Company's Published Efficiency Claims
Cognition says it evaluated Fusion with Fable 5.1 and GPT-6 Astra paired with SWE-2 as the sidekick across several coding agent benchmarks in partnership with Artificial Analysis and Vals AI. The company describes significant cost savings while maintaining frontier performance.
These are Cognition's self-reported evaluation results, not independently reproduced figures. In the same vein, the company claims Fusion is up to 39% more efficient than other model harnesses across major coding benchmarks. According to Nader Dabit's post, Fusion has also become the first multi-model coding agent listed on the Artificial Analysis Coding Agent Index.
How to Use It and Current Limitations
Fusion is used through model selection inside Devin Desktop and CLI: choose a Lead model and a Sidekick model, then adjust execution speed (Normal/Fast) and reasoning level. No separate install command or external integration procedure has been disclosed, so the in-app selection is the confirmed usage path.
The currently confirmed scope is as follows.
- Devin Desktop and CLI: the confirmed availability scope is Devin Desktop and CLI.
- Web app undecided: availability in the web app has not been confirmed.
A social reply raised a complaint suggesting personally supplied API keys (BYOK) cannot be used, but that is a third-party comment and cannot be confirmed as an official Cognition specification. Teams finding frontier-model-only usage too expensive can start by keeping a frontier model such as Fable 5.1 as Lead and delegating execution to a cost-effective model such as SWE-2. Since actual savings depend on the delegation setup, it is worth deciding upfront how much review scope the Lead should own in a real project.
Sources
- Cognition: Introducing Fusion in Devin Desktop & CLI
- Nader Dabit (@dabit3): Fusion announcement post
- Cognition: Devin Fusion