Cognition announced Fusion on September 11, 2026, bringing a two-agent coding harness to Devin Desktop and Devin CLI. A frontier lead model plans the work, handles ambiguity and reviews the result; a lower-cost sidekick takes on bounded engineering tasks such as exploring a repository, writing code and running tests. The pitch is straightforward: spend the expensive reasoning where it matters, and let the cheaper model do more of the legwork.
What Cognition Fusion is
Fusion is not a standalone model, editor or subscription. It is a multi-model harness inside Devin. Cognition recommends pairing Fable 5.1 as the lead with SWE-2 as the sidekick, while its benchmark material also reports a pairing built around GPT-6 Astra.
The user interacts with the lead. That agent creates a plan, interprets unclear requirements, decides what to delegate and reviews the returned work. The sidekick receives a bounded brief containing the task, constraints and success criteria, then explores the code, implements changes, runs tests and reports what happened.
That division matters because Fusion is not simply a router that picks one model before a task starts. Cognition describes a workflow in which responsibility can move back to the lead when the delegated work needs more judgment.
How the lead and sidekick work together
The two agents keep separate persistent contexts and tool access. They exchange briefs, results and feedback rather than continuously passing their complete conversation histories back and forth.
In practical terms, the lead does not need to replay every tool call to the sidekick. It can send the relevant assignment, receive the implementation and test results, and then review the outcome. This architecture is also designed to preserve prompt-cache benefits when the system changes models during a task.
The boundary is important: a sidekick is not being presented as a universal replacement for the frontier model. Cognition’s earlier Fusion material describes quality losses when a difficult task requiring subtle judgment is delegated poorly, while more mechanical work can cost less without losing quality—or can score higher in some tests. The useful question is therefore not “Which model wins everything?” but “Which part of this task needs the expensive model?”
What the benchmark numbers show
Cognition’s official FrontierCode 1.1 Extended comparison keeps each lead model paired with its corresponding solo baseline. The figures below are reported evaluation scores and average cost per task, not subscription prices or guaranteed customer bills.
| Lead configuration | Fusion configuration | Score | Average cost per task | Reported cost change |
| Fable 5.1 alone | Fable 5.1 + SWE-2 Fusion | 63.5 with Fusion; 63.6 alone | $1.67 with Fusion; $2.68 alone | 38% lower with Fusion |
| GPT-6 Astra alone | GPT-6 Astra + SWE-2 Fusion | 63.4 with Fusion; 63.1 alone | $2.34 with Fusion; $2.62 alone | 11% lower with Fusion |
The Fable pairing is the sharper cost example: the Fusion configuration comes within 0.1 point of the solo score while reducing the reported average task cost from $2.68 to $1.67. The Astra pairing moves in the other direction on score, reaching 63.4 instead of 63.1 while reducing the reported cost by $0.28 per task.
Those results are not universal parity. Across Cognition’s wider benchmark table, reported cost reductions range from 11% to 46%, and several Fusion configurations score below their corresponding solo models. On DeepSWE 1.1, for example, Fable 5.1 with Fusion scores 63.1 versus 64.3 alone, while Astra with Fusion scores 67.3 versus 67.6 alone. On Terminal-Bench 4, both the Fable and Astra Fusion configurations score below their solo baselines.
The Devin session display adds a separate example: it shows a 39% cheaper estimate than using Claude Fable 5.1 alone, with 113,000 tokens assigned to Claude Fable 5.1 and 97,000 to SWE-2, for 210,000 tokens total. The display labels that figure as an estimate based on pricing the sidekick tokens at the listed lead-model rates. It is not a controlled rerun of the same task with the lead model alone, so it should not be read as a guaranteed saving.
Are those dollar figures subscription prices?
No. The $1.67, $2.34 and other figures above are reported benchmark costs per task. They are not Devin plan prices and do not predict a customer’s bill for a particular repository.
Fusion is presented as a feature within Devin rather than a separate subscription. Official Devin documentation lists access on paid plans and excludes free and trial tiers. The documented minimum versions are Devin CLI 3000.10.20+ and Devin Desktop 3.10.0+.
Who should care about Fusion?
Fusion is most relevant if you already use Devin for software-engineering work and want a way to divide planning from implementation. The design may be attractive for repository exploration, routine code changes and test runs that can be described with clear constraints. Tasks dominated by architectural judgment or ambiguous requirements still depend heavily on the lead’s review and its decision about what to delegate.
For developers evaluating the product, the important checklist is short:
- Treat benchmark scores as task-specific measurements, not a universal quality rating.
- Compare the same lead model with and without Fusion.
- Keep average cost per task separate from subscription pricing.
- Look at both sides of the result: lower cost can arrive with a slightly lower score on some benchmarks and a higher score on others.
Fusion’s real change is architectural rather than magical. It gives Devin a lead agent that can reason about the job and a cheaper agent that can execute defined pieces of it. The reported results suggest that trade-off can reduce evaluation cost, but the mixed scores make the final decision task-dependent.
