Cognition Launches Devin Fusion, Scoring 61.7 on Coding Agent Index

On September 12, Cognition released Devin Fusion, a coding agent harness billed as "the most efficient frontier coding agent today." Its core architecture pairs a frontier main model with a low-cost sidekick model: a strong user-preferred model handles planning, while a cheaper model executes, with base models configurable (e.g., Claude Fable 5.1 / GPT-6). The company also launched Fusion in Devin CLI, claiming that as a harness for the Fable and Astra models it is 39% cheaper than prior approaches on coding benchmarks.

Confirmed

• Artificial Analysis independently evaluated Devin Fusion on launch day—the first multi-model coding agent to enter its Coding Agent Index (v1.5).
• The Fable 5.1 (xhigh) + SWE-2 (medium) configuration scored 61.7, cutting per-task cost by 36% while scoring only about 4% below Claude Code.
• Artificial Analysis published full per-category results for two configurations.

Why it matters

• Devin Fusion is the first proof that a "strong model plans, cheap model executes" multi-model architecture can compete with top-tier coding agents—trailing Claude Code only slightly on performance while slashing costs, directly shaking up the price-performance landscape of coding agents.

2026-09-12 ~ 2026-09-12 · 5 related posts

• [source] Cognition's Devin CLI Fusion Cuts Coding Benchmark Costs 39% by Splitting Planner and Executor — DeryaTR_ · 2026-09-12
• [source] Devin Fusion benchmarks: frontier lead model plus cheap sidekick cuts coding agent costs — ArtificialAnlys · 2026-09-12
• Devin Fusion scores 61.7 on Coding Agent Index, nearly matching Claude Code at 36% less cost — ArtificialAnlys · 2026-09-12
• [source] Devin Fusion hits 61.7 on Coding Agent Index, costs 36% less than Claude Code per task — ArtificialAnlys · 2026-09-12
• Devin Fusion Debuts: Multi-Model Coding Agent Scores 62, First on AA Index — AccBalanced · 2026-09-12