The usual attempt to feed an entire codebase to a single heavy language model quickly hits a wall of token bills. On September 11, Cognition announced the availability of Fusion in Devin Desktop and CLI, bringing a dual-model setup from the cloud directly into local developer tools. The mechanics are intensely pragmatic: instead of one monolithic model, a lead agent and an executor work in tandem. The lead handles planning and communication, while the executor writes code, fixes bugs, and runs tests.
The Economics of Two Roles
Cognition recommends using combinations like Claude paired with specialized executors. In tests, this setup scores 61.7 points compared to 62.2 for monolithic configurations, while cutting costs by 36%. A similar picture emerges on other benchmarks: alternative configurations score 58.9 versus 61.6 while reducing costs by 39%.
A dual-agent architecture saves up to 40% of the budget by delegating routine code generation to an executor, maintaining quality on par with a single heavy model.
However, there is no universal formula, and savings inevitably come with a trade-off in quality. For instance, on Terminal-Bench the cost drops by 40%, but the score falls from 55.6 to 50.0 points. Cognition rightly emphasizes that you cannot blindly apply these numbers to an arbitrary working project: benchmark measurements are tightly bound to specific configurations.
The Expensive Token Paradox
A more expensive model per token sometimes turns out cheaper overall thanks to efficient planning: it hands off tasks to the executor faster, duplicates less context, and ultimately burns fewer resources per task. Division of labor works even within algorithms. The winner is not the one who hoops up the cheapest tokens, but the one whose planner knows how to delegate work on time without turning the local environment into an expensive query dump.