Cognition launches Fusion in Devin: a two-model pairing cuts task costs by up to 39%

Cognition has added a Fusion mode to Devin Desktop and the CLI, in which two models work on a single task at once. A frontier model acts as the "lead": it plans, resolves ambiguities, and verifies the result. A cheaper model, the "sidekick," explores the code, makes edits, and runs tests. According to the company, with the Fable 5.1 + SWE-2 pairing, a task costs 36% less than with Claude Code on Fable 5.1 ($7.90 versus $12.36 in a normalized comparison), and the final score is nearly the same: 61.7 versus 62.2. For GPT-6 Astra paired with SWE-2, the savings are 39% relative to Codex ($4.54 versus $7.47), with a score of 58.9 versus 61.6.

Why not ordinary routing

The classic idea is to give simple tasks to a cheap model and hard ones to an expensive model. Cognition explains why this works poorly. You can't judge difficulty from the first prompt: "fix the bug" might mean one line or a re-architecture of the product, and that only becomes clear after reading the code. On top of that, switching models mid-task breaks the prompt cache, and reprocessing the context is then billed at the expensive model's rate.

Fusion avoids both problems. Two parallel agents run, each with its own persistent context and set of tools. The full conversation is not passed between them, only briefs, results, and feedback. The lead writes the sidekick a task with constraints and success criteria, and does not get bogged down in intermediate tool output. Both agents keep their cache. The user always talks to the frontier model, which can take the work back at any point if the sidekick isn't coping.

Benchmark numbers

Together with Artificial Analysis and Vals AI, Cognition compared Fusion with single models. Scores and dollar costs for individual tests:

• DeepSWE 1.1: Fable 5.1 — 64.3 ($14.63); Fusion on Fable — 63.1 ($7.88, −46%). Astra — 67.6 ($7.88); Fusion on Astra — 67.3 ($4.69, −40%).
• Terminal-Bench 4: Fable — 57.6 ($17.46); Fusion — 56.1 ($13.37, −23%). Astra — 55.6 ($10.08); Fusion — 50.0 ($6.06, −40%).
• SWE-Atlas QnA: Fusion on Fable scores 65.9 versus 64.8 for the single model, at a cost of $5.00 instead of $7.57 (−34%).
• Vals Code Migration: Fusion on Fable — 57.3 versus 54.6 at −41% cost; for Astra the cost is down 20%, but the score drops from 67.7 to 61.3.
• FrontierCode 1.1 Extended: for Fable, quality is 63.5 versus 63.6 at −38%; for Astra, the savings are only 11%.

The picture is uneven. On Terminal-Bench and Code Migration, the Astra + SWE-2 pairing shows a noticeable quality drop, so "frontier-level on every task" is an assumption to treat with caution. Note also that the data is published by Cognition itself, albeit with third-party evaluators involved.

Expensive models can make the system cheaper

The most interesting finding concerns cost per task rather than cost per token. When Cognition replaced Opus 4.8 with Fable 5 as the lead, sessions became 9% cheaper on average despite a per-token price twice as high, and the FrontierCode score rose. Fable delegated earlier and wrote more precise briefs, while Opus micromanaged the sidekick and redid its work.

The same holds for the executor. SWE-2 costs $0.75 per million tokens versus $0.20 for GPT-5.6 Luna (+275%), yet the total cost on FrontierCode barely changes: $2.34 versus $2.39, and the score is higher — 63.4 versus 62.0. A strong sidekick takes fewer turns and needs fewer review rounds.

The harness, however, has to be tuned for each pair. A weak sidekick needs more detailed briefs, is better not allowed to argue with the plan, and shouldn't be handed research that affects the plan. A strong executor can be left to handle implementation details and given the initial code reconnaissance.

What to try

You can install Devin CLI with the command curl -fsSL https://cli.devin.ai/install.sh | bash. In Fusion, you select two models: Cognition recommends Fable 5.1 as the lead and SWE-2 as the sidekick. If you track agent spending, it makes sense to run two or three pairings on your own tasks and compare not price lists but the final cost of a completed task. This is especially relevant for routine migrations and large codebases, where the sidekick does most of the work.