Cognition’s Fusion uses a frontier lead and cheaper sidekick to cut coding-agent task cost

Cognition released Fusion in Devin Desktop and Devin CLI on September 11, 2026. Rather than running one model for an entire coding task, Fusion keeps two persistent agents: a stronger frontier “lead” handles planning and review, while a cheaper “sidekick” performs much of the implementation work. The agents maintain separate contexts and exchange briefs, results and feedback, which Cognition says preserves prompt-cache efficiency while moving a large share of execution tokens onto the cheaper model. Cognition recommends using Fable 5.1 as the lead with its SWE-2 model as the sidekick.

Coding-agent economics are often discussed as model price per million tokens, but a harness can change which model consumes those tokens and how much rework a task needs. Fusion makes that orchestration strategy a product feature. In Cognition’s published tests, the two-model setup usually cut total task cost substantially while keeping benchmark scores near the single-frontier-model baseline, although several workloads lost measurable accuracy. For teams running agents heavily, that makes harness design and task-level cost as important as nominal model pricing. The trade-off is that these are Cognition-run benchmarks rather than independent production measurements.

Fusion assigns planning, review and higher-level decision-making to the lead model while the sidekick handles more of the execution. Both agents persist through the task instead of invoking a cheap model as a stateless one-off tool, so each can accumulate its own working context.

Across Cognition’s published benchmark table, Fusion reduces reported task cost by roughly 11% to 46% depending on the lead model and benchmark. Some scores remain almost unchanged or improve slightly, while others fall—for example Astra on Terminal-Bench 4 drops from 55.6 to 50.0 and on the Vals Code Migration benchmark from 67.7 to 61.3. That makes the result a cost-quality trade rather than a universal efficiency win.

Cognition argues that the cheapest model per token is not necessarily cheapest per solved task. Its FrontierCode example reports Astra Fusion with SWE-2 at $2.34 and a 63.4 score versus a cheaper-per-token Luna sidekick at $2.39 and a 62.0 score. The useful metric therefore becomes total cost for acceptable task completion, not the model’s list token price in isolation.

The lead and sidekick maintain distinct contexts and communicate through explicit briefs, results and feedback. Cognition says this allows each role to keep a coherent prompt history and prompt cache rather than repeatedly moving a whole large context between models.

The benchmark suite, configurations and cost calculations are published by Cognition, which also sells Devin and provides SWE-2. They are useful engineering evidence but not neutral measurement. Production teams should test their own task mix, especially workloads where a small quality loss can create expensive review or retry loops.