Cognition’s Fusion cuts coding-agent costs with a two-model workflow

Cognition has added Fusion to Devin Desktop and the Devin CLI, where the mode runs two cooperating coding agents. A frontier-class lead model plans and reviews the work, while a lower-cost sidekick handles delegated tasks. Cognition recommends Fable 5.1 as the lead and SWE-2 as the sidekick, with Astra available as an alternative lead. The company reports cost reductions of up to 39% across major coding-agent benchmarks compared with running the lead model alone.

Delegation keeps both caches warm

Conventional model routers often choose a cheap or expensive model from the initial request, before either model has inspected the repository. A prompt such as “fix xyz bug” might require a one-line edge-case fix or a broad architectural change. Providers can cache an unchanged conversation prefix to reduce repeated processing costs, but transferring a session between models can forfeit much of that reuse.

Fusion keeps both agents active, each with its own context window, prompt cache, and tools. The lead owns the session and delegates bounded pieces of work to the sidekick. The agents exchange briefs, results, and review feedback without copying their full conversations between contexts.

A typical session starts with the lead exploring the codebase and preparing a plan. It gives the sidekick constraints, relevant context, and success criteria, then reviews the returned implementation. The lead can request corrections, integrate the work, or continue the task itself.

Costs fall, scores vary

Cognition partnered with Artificial Analysis and Vals AI to compare Fusion with solo-model runs across repository work, terminal tasks, codebase questions, and migrations. Across the ten published comparisons, reported costs fell by 11% to 46%. Score changes ranged from a 2.7-point improvement to a 6.4-point decline. Higher scores are better, and costs are reported in US dollars.

The Fable-led pairing produced the steadier quality profile. Four benchmark scores moved by no more than 1.5 points, while the migration score improved by 2.7 points. With Astra leading, DeepSWE and FrontierCode remained nearly flat, but Terminal-Bench fell by 5.6 points and Vals Code Migration fell by 6.4 points.

The results show that savings depend on both the model pair and the workload. Benchmark averages provide a controlled comparison, while production adoption requires tests against representative repositories, tools, and acceptance criteria.

Stronger models can lower the bill

A sidekick with a higher token price can reduce total session cost when it reaches a correct implementation in fewer turns. Better first attempts also reduce the lead model’s review, correction, and rework. The resulting session can use fewer tokens even when each token costs more.

Cognition observed the same effect in a separate lead-model experiment using Fable 5 and Opus 4.8. Fable’s nominal per-token price was twice Opus 4.8’s, while Fable-led sessions cost 9% less on average and scored higher on FrontierCode. Cognition attributes the result to Fable delegating earlier and writing clearer briefs; Opus spent more turns directing the sidekick and redoing its work. The released recommendation names Fable 5.1, while this experiment used Fable 5.

Price per completed task captures these interactions more accurately than token price alone. For agent systems, the harness influences how much exploration, delegation, review, and correction each model performs.

Each pairing changes the playbook

Fusion uses pair-specific harness policies because model capability affects the amount of instruction, autonomy, and review that produces the best result. Cognition tunes three main controls:

• Brief specificity: Less capable sidekicks receive prescriptive instructions, constraints, and implementation details. Stronger models receive more discretion.
• Pushback: Stronger sidekicks can challenge the lead’s plan and flag mistakes. Cognition found that encouraging the same behavior from weaker sidekicks increased cost and reduced performance.
• Code exploration: The lead handles planning-time repository exploration when paired with a weaker sidekick. Stronger sidekicks can help inspect the codebase before implementation begins.

Results from one combination therefore do not transfer automatically to another. A new lead or sidekick requires its own delegation policy and benchmark pass.

Install, then test on real work

Fusion is available in Devin Desktop and Devin CLI. CLI users can install the client with Cognition’s shell command:

curl -fsSL https://cli.devin.ai/install.sh | bash

The command downloads and executes a remote script, so managed environments can fetch and inspect the script before running it. After installation, start a session, select Fusion, and choose a supported lead and sidekick.

A production trial can compare Fusion with solo-model sessions using the following measurements:

• Task completion rate and automated test results
• Total session cost, including review and correction turns
• Wall-clock time and the number of delegation rounds
• Consistency across repeated runs of similar tasks
• Performance by workload, especially migrations, terminal-heavy tasks, and ambiguous bug reports

The published results support substantial savings for several workloads, particularly with Fable 5.1 and SWE-2. Astra’s larger regressions on terminal and migration benchmarks show why each repository and task mix needs its own quality threshold before broader deployment.