Windsurf: Devin Fusion Comes to Desktop and CLI
WindsurfWindsurf's Devin Desktop and Devin CLI gained Fusion, the dual-model harness that previously ran only in Devin Cloud, letting developers pair a frontier "lead" model for planning and review with a cheaper "sidekick" model for execution. Cognition recommends pairing Claude Fable 5.1 or GPT-6 Astra with its own new SWE-2 model, citing cost reductions of roughly 36-40% on the Artificial Analysis Coding Agent Index versus using a single frontier model alone. Fusion requires a paid plan on Devin CLI 3000.10.20 or Devin Desktop 3.10.0 or later, and each model in a pairing is billed separately at its own rate.
Key Takeaways
- Fusion is no longer Cloud-only; the dual-model harness Cognition built for Devin Cloud months ago now runs locally inside Devin Desktop and the Devin CLI.
- Choosing Fusion means picking two models instead of one: a frontier "lead" that plans, interprets ambiguity, and reviews work, paired with a cheaper "sidekick" that explores code, implements changes, and runs tests.
- Cognition's own numbers show Fable 5.1 plus SWE-2 cuts cost by 36% and Astra plus SWE-2 by 39% on the Artificial Analysis Coding Agent Index versus running Claude Code or Codex alone.
- The lead and sidekick agents exchange only task briefs, results, and feedback rather than full conversation histories, which Cognition says enables independent prompt caching and helps control cost.
- The cost savings are not uniform across tasks; on FrontierCode 1.1 the Astra-plus-SWE-2 pairing saves only about 11%, while on DeepSWE 1.1 it saves around 40%.
- Fusion is limited to paid plans on Devin CLI 3000.10.20+ or Devin Desktop 3.10.0+, with each model in a pairing billed separately at its own per-token rate.
Sources & Mentions
2 external resources covering this update
Fusion Leaves the Cloud
Cognition introduced Fusion for Devin Cloud months earlier in 2026 as a way to cut the cost of agentic coding without giving up frontier-level results. On September 11, Windsurf's Devin Desktop and Devin CLI gained the same capability, letting developers run the hybrid harness locally instead of only through Devin Cloud sessions.
How Fusion Works
Choosing Fusion means picking two models instead of one. The frontier "lead" model owns the task: it plans the work, interprets ambiguous instructions, and reviews everything before it ships. A cheaper "sidekick" model handles the mechanical side of the job, exploring the codebase, implementing changes, and running tests. Rather than sharing full conversation histories, the two agents only exchange task briefs, results, and feedback, which Cognition says enables independent prompt caching for each model and keeps the overall bill down. For best results, Cognition recommends pairing Claude Fable 5.1 or GPT-6 Astra as the lead with its own SWE-2 model as the sidekick.
The Cost Case
On the Artificial Analysis Coding Agent Index 1.5, Cognition reports that Fusion pairing Fable 5.1 with SWE-2 costs 36% less than running Claude Code alone, and pairing Astra with SWE-2 costs 39% less than running Codex alone. The savings are not uniform across every benchmark, though: on Cognition's own FrontierCode 1.1 test, the Astra-plus-SWE-2 pairing saves only about 11%, while on DeepSWE 1.1 the same pairing saves around 40%. Cognition's broader argument is that price per token is only part of the story, since frontier models tend to be more token-efficient and need fewer back-and-forth exchanges to finish a task, meaning a pricier model can sometimes still work out cheaper overall.
Getting Started
Fusion is restricted to paid Devin plans running Devin CLI 3000.10.20 or later, or Devin Desktop 3.10.0 or later; it is not included on free or trial tiers. Selecting Fusion happens through a visual model picker in the CLI, where users choose their lead and sidekick separately, and each model in the pairing is billed at its own per-token rate rather than a single blended price.