On September 11, 2026, Cognition shipped Fusion to Devin CLI and Devin Desktop, officially billing it as "39% cheaper across coding benchmarks." The next morning, Artificial Analysis published an independent evaluation: the first multi-model coding agent to appear on its Coding Agent Index. The Fable 5.1 + SWE-2 configuration scored 61.7 versus 62.2 for Fable 5.1 running in Claude Code, at $7.90 versus $12.40 per task. The $4.50 difference sounds trivial, but a team running 100,000 tasks a year saves $450,000.
Where the Money Goes
Coding agent spending burns in two places: input (the agent repeatedly feeds whole-repo context to the model) and output (reasoning, code, tests, and error-retry loops are all billed per token). The industry's approach over the past two years was to make models bigger so they get things right in one pass. That works, but marginal returns are shrinking. Cognition changed direction: instead of changing the model, change the division of labor.
Fusion's structure is "a lead plus a sidekick." The lead uses a frontier model (Fable 5.1 or GPT-6 Astra, both at xhigh) for session planning, disambiguating instructions, and final review. The sidekick uses the cheaper SWE-2 (medium) for codebase exploration, implementation, and running tests, then reports back. The two agents hold independent contexts and exchange three things: a task brief with constraints and success criteria, the delegated work's results, and the lead's review feedback. The frontier model retains control throughout and can reclaim work when the sidekick's task gets too hard.
Why Now
Three years ago this made no sense: the capability gap meant retries on cheap models cost more than the savings. By 2026, mid-tier models handle mechanical work (exploration, boilerplate, tests, reporting errors) well enough while frontier prices keep climbing — the price curves have pulled apart, leaving room for a routing layer. A second engineering driver: agent sessions now span hundreds of thousands of tokens, so input costs dominate. Cutting output tokens alone is no longer enough; you have to restructure the input side.
The Real Engineering Is in the Cache
Fusion is easy to misread as request-level routing. It is not: request-level routing picks one model per request; Fusion runs two models simultaneously on one task. The unavoidable pitfall is that switching models mid-session invalidates prompt caches — the accumulated KV cache is lost and the new model must do a full prefill, a "cache tax" that can eat the model price difference. Fusion's solution: each agent holds its own independently cacheable context rather than sharing one ever-growing transcript. Cognition has also described a complementary trick from earlier work: swap models at context-compaction points, where the cache is invalidated anyway.
The Numbers: Nearly Identical Scores, 36% Cheaper
Artificial Analysis used Coding Agent Index v1.5:
| Configuration | Index Score | Cost per Task | Time per Task | |---|---|---|---| | Fable 5.1 xhigh + SWE-2 medium (Fusion) | 61.7 | $7.90 | 35.8 min | | Fable 5.1 max with fallback (Claude Code) | 62.2 | $12.40 | 34.8 min | | Astra xhigh + SWE-2 medium (Fusion) | 59 | 43% cheaper | 31% faster |
A 0.5-point gap, 36% cheaper, one minute slower. Both configurations sit on the score-versus-cost Pareto frontier. But the composite hides asymmetry: on SWE-Atlas QnA the Fusion config scored 65.9 vs 64.8 (better), while DeepSWE 1.1 dropped to 63.1 vs 64.3 and Terminal-Bench 4.0 to 56.1 vs 57.6. Retrieval-heavy repo Q&A improved; judgment-heavy work lost points.
Cognition also published more extreme figures: pairing Astra-led sessions with a $0.75/M-token SWE-2 flank scored 63.4 at $2.34; swapping to $0.20/M Luna scored 62.0 at $2.39 — one-sixth the price yet slightly more expensive and lower-scoring. Cheaper flanks are not always better.
The Same Day, Sakana Made It Theory
Sakana AI released Fugu Max and Fugu Ultra v2 on September 11 with a blunter framing: a decade of single-axis competition made models bigger and pricier, but the real frontier is two-dimensional — capability and cost. Running a simple data query on a multi-trillion-parameter model is waste. Fugu Max is priced at $2/M input and $6/M output tokens, claimed 40–60% cheaper on output than Sonnet 5, GPT 5.6 Terra, and Kimi K3, pushing the cost-performance Pareto frontier outward on seven of ten benchmarks. Fugu Ultra v2 goes the other way: 48.3 on Chartography (vs reported 27.3 for Opus 5 and 29.5 for Fable 5), 74.3 on DeepSWE, outperforming models three to five times pricier per token — notably without Fable 5, Fable 5.1, or GPT-6 Astra in its agent pool.
| System | How it splits | Basis | |---|---|---| | Devin Fusion | By role: planning to lead, execution to sidekick | In-session role division | | Sakana Fugu Max | By task: route to "cheapest model that can finish" | Task difficulty estimation | | Copilot HydraFusion | By dimension: four-axis scoring picks execution workflow | Four-dimensional prompt feature scoring |
The last row comes from a Microsoft engineer's walkthrough of the GitHub Copilot HydraFusion engine: each prompt is scored on reasoning depth, code-generation complexity, debugging difficulty, and tool-orchestration needs, producing a HyDRA score that selects one of three execution workflows, with models assigned per stage. Unlike single-model end-to-end Auto mode, HydraFusion decomposes by stage.
One Underlying Judgment, Three Implementations
All three reflect the same insight: different parts of a task have different "difficulty density," and difficulty density need not be tied to price per token. Planning demands understanding ambiguous requirements and weighing tradeoffs — premium models earn their cost there. Exploring codebases, patching boilerplate, running tests, and pasting errors back is mid-tier-model work; using a frontier model pays a premium for repetitive labor.
Three Caveats
1. Fusion's cost figures come from benchmarks, not your own PRs. Cognition's claim that 88% of its own engineering team's merged PRs used Fusion's auto-routing is also a vendor statement without third-party review. 2. Sakana's pricing and benchmarks are self-reported. Fugu Ultra v2's training cutoff is August 28, 2026, its full agent pool is undisclosed, and SWEFish is its own internal benchmark. None of the three datasets has been independently replicated. 3. Score asymmetry matters. SWE-Atlas QnA up, Terminal-Bench 4.0 down suggests the judgment-heavy portions of delegated work lose points. How much you save depends on your task mix: exploration-and-testing-heavy tasks route well; judgment-heavy tasks save less and may lose quality.
Things to watch: whether Cognition's 88% PR routing rate gets an independent audit, and whether Artificial Analysis makes multi-model configurations a permanent category — it explicitly noted this was the first multi-model coding agent in the index.
Sources
1. Artificial Analysis independent evaluation of Devin Fusion (Coding Agent Index v1.5, 2026-09-12), relayed via AI Primer and AGI Hunt. 2. Cognition official announcement and Fusion technical notes (2026-09-11/12): lead/sidekick division, three handoff artifacts, 39% cross-benchmark cost reduction, $0.75/M vs $0.20/M flank comparison. 3. Sakana AI, "Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier" (2026-09-11). 4. Microsoft engineer walkthrough of GitHub Copilot HydraFusion (2026-09-12). 5. ExplainX, "Devin Fusion CLI: 39% Cheaper Coding Agent Runs" (2026-09).