I’ve been using Devin in the cloud for quite a while, but Fusion in Devin Desktop and CLI has changed how I use it locally.
I can now pair Fable 5.1 or GPT-6 Astra as the lead model with SWE-2 as the sidekick. Across several serious coding tasks, both combinations have performed very well for me. The lead model handles planning, ambiguity, and review, while SWE-2 takes on more of the execution work.
The biggest difference is efficiency. Before Fusion, just a few demanding tasks run entirely on Fable could noticeably drain my weekly allowance. With the $200/month Devin Max plan, I can now use Fusion across multiple pieces of work much more comfortably, and my weekly quota lasts significantly longer. So far, the quality has remained strong enough that this does not feel like a simple capability-for-cost trade-off.
This is a personal observation from real projects, not a controlled benchmark, but the improvement has been meaningful.
Has anyone else tested Fable 5.1 + SWE-2 or GPT-6 Astra + SWE-2 in Devin Fusion? How has it performed on larger real-world tasks, and how do its reliability and quota consumption compare with running the frontier model alone?
