The Pulse
Cognition Brings Fusion’s Two-Model Coding System to Devin Desktop and CLI
Cognition has released Fusion in Devin Desktop and Devin CLI, pairing a frontier “lead” model with an execution-focused sidekick. The company says the system cuts task costs by as much as 46% on selected coding benchmarks while retaining si
AI.info Team ·
46% is the largest per-task cost reduction Cognition reports for Fusion, its two-model coding harness, on the DeepSWE 1.1 benchmark. The figure comes from a configuration pairing Fable 5.1 as the lead model with Cognition’s SWE-2 as a sidekick: the Fusion run scored 63.1 versus 64.3 for Fable alone, while the average task cost fell from $14.63 to $7.88. Cognition released Fusion for Devin Desktop and Devin CLI on September 11, moving an architecture that had been available in its cloud product onto developers’ local workstations.
Fusion is not a new foundation model. It is a system for coordinating two agents with different jobs: a more capable, higher-priced model plans work, resolves ambiguity and reviews output, while a lower-cost model investigates code, writes changes, runs tests and returns results. Cognition’s premise is that coding agents do not need frontier-grade reasoning on every turn, but that handing an entire task to a cheaper model risks losing the judgment needed for a production change.
The release arrives a day after Cognition introduced SWE-2, the coding model it recommends as Fusion’s execution sidekick. Together, the launches show Cognition placing its product bet not only on better individual coding models, but also on the software layer that decides how those models divide work. For engineering teams paying by token or by completed agent task, the result could matter as much as a modest increase in benchmark accuracy.
Fusion splits planning from execution
Devin users selecting Fusion choose two models rather than one. Cognition calls the frontier model the lead and the lower-cost partner the sidekick, recommending Fable 5.1 and SWE-2 as the preferred combination. The lead controls the session, makes the plan, interprets unclear requirements and reviews the result; the sidekick receives a focused brief with constraints and success criteria.
That division avoids a common weakness in basic model routing, according to Cognition. A prompt such as a request to fix a bug does not reveal whether the work requires a one-line patch or a redesign across a codebase. Difficulty also changes after an agent starts inspecting repositories, reproducing failures and running tests, so a one-time decision to send a task to a cheap model can fail after the work becomes more complex.
Cognition instead keeps both agents running in parallel with separate persistent contexts and tool access. Rather than transferring the full conversation and every tool output between models, the agents exchange short briefs, results and feedback. The company says that design preserves prompt-cache benefits, because each model continues working from its own context rather than rebuilding a long history after every handoff.
The arrangement also gives the lead a continuing supervisory role. A sidekick can implement a scoped change or handle repetitive test and repository work, but the lead can identify a flawed approach, request corrections or take over. Cognition argues that oversight is the distinction between Fusion and a router that simply assigns a model to an entire task at the start.
Benchmark tables show lower costs, not uniform gains
Cognition worked with Artificial Analysis and Vals AI to evaluate Fusion across five coding-agent benchmarks. The company presents results for Fable 5.1 and GPT-6 Astra as lead models, each paired with SWE-2. Costs decline in every published comparison, but benchmark scores do not move in one direction across all workloads.
On DeepSWE 1.1, Fusion with Fable 5.1 cut the listed task cost 46%, from $14.63 to $7.88, with a 1.2-point score decrease. Fusion with Astra reduced cost 40%, from $7.88 to $4.69, while the score moved from 67.6 to 67.3. On Terminal-Bench 4, the Fable pairing saved 23% and the Astra pairing saved 40%, though both recorded lower scores than their single-model baselines.
Other results show Fusion matching or exceeding the lead model’s score. The Fable-SWE-2 pairing recorded 65.9 on SWE-Atlas QnA versus 64.8 for Fable alone, while listed cost dropped 34%, from $7.57 to $5.00. On Vals Code Migration, Fusion scored 57.3 versus 54.6 for Fable by itself and reduced cost 41%, from $70.97 to $42.00.
The table’s widest score gap appears in a different direction: Astra alone scored 67.7 on Vals Code Migration, while Astra paired with SWE-2 scored 61.3, even as cost fell 20%. Those figures make the limits of a single top-line efficiency claim clear. Fusion’s published results suggest that the system can save money while holding close to the original model on many tasks, but performance depends on the benchmark, lead model and the work delegated to the sidekick.
Prompt caching becomes part of the cost equation
Cognition’s technical case for Fusion rests heavily on prompt caching. Long-running coding sessions can accumulate large repositories, terminal outputs, test logs and planning traces in their working context. Switching a task from one model to another can force the newly selected model to process material again, turning a cheap-routing decision into an expensive cache miss.
The company says Fusion avoids that penalty by giving each agent its own ongoing context from the beginning. A lead does not need every command the sidekick ran in order to judge a completed patch, while the sidekick does not need the lead’s entire planning history to edit a defined portion of a repository. The compressed exchange between agents is meant to reduce duplicated reading while preserving enough information for review.
Cognition also says that a more expensive model can lower total cost when it makes fewer mistakes or provides clearer instructions. In one comparison, replacing Opus 4.8 with Fable 5 as the lead reduced average session cost by 9%, despite Fable’s higher stated per-token price. Cognition attributes the outcome to earlier delegation and better briefs from Fable, which reduced repeated work by the sidekick and review work by the lead.
The same argument shapes the choice of SWE-2 over a cheaper execution model. Cognition lists GPT-5.6 Luna at $0.20 per million tokens and SWE-2 at $0.75 per million tokens, a 275% increase in list price. Yet in an Astra-led Fusion configuration on FrontierCode 1.1, the SWE-2 version scored 63.4 at $2.34 per task, compared with Luna’s 62.0 at $2.39, according to the company’s data.
Each model pair needs its own operating rules
Fusion does not treat the lead and sidekick as interchangeable components. Cognition says it tunes the harness for particular model pairings, including how detailed a lead’s assignment should be, whether a sidekick should challenge the plan, and which exploration work belongs with the lead. A weaker sidekick may need highly prescriptive instructions, while a stronger one can investigate implementation details and challenge a mistaken assumption.
Delegation remains risky when judgment itself is the deliverable. In an earlier Fusion analysis, Cognition described a hard React and Redux feature in which the sidekick handled coding but missed subtle product intent; the task became cheaper but the score fell sharply. A separate task involving a difficult Java integration performed better with the sidekick because the work was mostly mechanical reuse of upstream code.
That distinction matters for users evaluating agents beyond leaderboard results. Mechanical migrations, broad deprecations, repetitive edits and long test runs may fit the sidekick model well. Ambiguous product features, architecture decisions and bug investigations that determine what should be changed may require the lead to remain closer to the work.
“The industry needs to move from maximizing usage metrics to maximizing outcomes,” Scott Wu, Cognition’s co-founder and chief executive, wrote in June when the company announced an AI productivity guarantee for enterprise customers.
Wu’s framing helps explain why Cognition emphasizes price per completed task rather than token price. Fusion asks buyers to judge an agent on whether it produces an acceptable pull request at a lower cost, not on whether it consumes fewer tokens in isolation. Benchmark costs remain estimates tied to defined tests, however, and teams will need to compare them with their own repositories, build systems and review standards.
Devin Desktop turns Fusion into a local workflow
Fusion had already operated in Devin Cloud, but the September 11 release makes it available in Devin Desktop and through Devin CLI. Devin Desktop is Cognition’s renamed successor to Windsurf, the agentic development environment Cognition acquired in 2025. Local availability puts the lead-and-sidekick system in the same environment where developers inspect diffs, run commands and decide whether to merge a change.
Cognition’s announcement presents a CLI session in which Fusion uses 113,000 tokens from Fable 5.1 and 97,000 from SWE-2, for 210,000 tokens in total, while claiming the run costs 39% less than Fable 5.1 alone. The company’s public benchmark table contains larger percentage savings on particular workloads, including the 46% DeepSWE result, so the 39% figure describes a separate illustrated session rather than a universal reduction.
The release leaves several practical questions for engineering organizations. How often will the lead detect a sidekick’s subtle but plausible error before code review? How much tuning will teams need for languages, frameworks and internal development practices that do not resemble public benchmarks? Whether the lead catches those failures on unfamiliar codebases is the question Fusion’s new Devin Desktop and CLI users will now answer.