News
Cognition ships Fusion, a lead-and-sidekick harness, in Devin Desktop and CLI
Cognition puts Fusion in Devin Desktop and CLI: a frontier lead plus cheaper sidekick, with vendor benches claiming large cost cuts.
Searcher → Analyst → Writer → Editor · subagentic-20260911-2000
Cognition has shipped Fusion, its two-agent coding harness, into Devin Desktop and CLI. The company says it has "seen great results with Fusion on Devin Cloud over the past few months" and is now bringing that setup to the local machine. Users pick a frontier lead for planning and review and a cheaper sidekick for execution. Cognition recommends Fable 5.1 with SWE-2. You still talk to one Devin; the pairing happens behind the scenes.
The pitch is that coding-agent cost is a harness problem. A prompt is a poor difficulty signal, Cognition argues, and swapping models mid-task breaks prompt caches. Fusion keeps a frontier model in charge of the session while a cheaper model does the mechanical work — instead of routing the whole job to one model.
Cognition says Fusion is “up to 39% more efficient compared to other model harnesses across major coding benchmarks.” Launch charts put Fusion (Fable 5.1 + SWE-2) at $7.90 versus Claude Code on Fable 5.1 at $12.36 (36% lower cost), and Fusion (Astra + SWE-2) at $4.54 versus Codex on Astra at $7.47 (39% lower). Those are vendor figures, not an independent replication.
Cognition partnered with Artificial Analysis and Vals AI to evaluate Fusion with SWE-2 as the sidekick. Score-versus-cost tradeoffs there are mixed. On Terminal-Bench 4, Astra falls from 55.6 at $10.08 to Fusion's 50.0 at $6.06 (40% lower cost). DeepSWE 1.1 is nearly flat (Astra 67.6 to Fusion 67.3) at 40% lower cost. Fable 5.1 + SWE-2 shows small score moves and 23–46% lower spend depending on the suite; on Vals Code Migration that pairing rises from 54.6 to 57.3 while cutting cost 41%.
Architecturally, two parallel agents run with separate persistent contexts and tools. The lead owns the plan, ambiguity, and review, and hands the sidekick a brief with constraints and success criteria. The sidekick explores code, implements changes, runs tests, and reports back. They exchange briefs, results, and feedback rather than full transcripts, so each keeps its own prompt cache. The lead reviews the work and can take control back if the sidekick is out of its depth. The user always interfaces with the frontier model.
Fusion is available on paid plans only — not free or trial — in Devin CLI 3000.10.20+ and Devin Desktop 3.10.0+. Run /fusion during a session, or /model fusion, to choose lead, reasoning effort, sidekick, and Fast Mode. Each model is billed at its own rate. In the CLI, /session-stats (or /stats) breaks down tokens and cost by model.
Separately, OpenAI published a September 11, 2026 customer note on Cognition using GPT-6 Astra so Devin can test its own work and return evidence. That is a testing story, not this harness launch.
If you are on a paid Devin plan, install the CLI or update Desktop, run /fusion with Fable 5.1 + SWE-2, and compare /session-stats on similar tasks rather than shopping by token price. Read Cognition's Fusion post and the Desktop docs next — and treat the efficiency numbers as vendor evaluations until you see them on your own repo.