Devin Fusion with an Opus 5.5 lead and GPT-6 Luna sidekick tops the Artificial Analysis Coding Agent Index at 68.1, at what Cognition says is 31% lower cost per task than Claude Code with Opus 5.5

Background

Coding-agent bills keep rising: frontier models are strong, but long tasks spend many tokens on reading code and running tests that don't need top intelligence. Simple difficulty routing fails because difficulty is often unknown until the code is read, and switching models mid-task breaks prompt caches. Cognition's Fusion, launched September 11, is a two-model harness built for this; its Artificial Analysis result was updated on October 19.

How it works

Fusion runs two agents with their own persistent context and tools. The lead, on a frontier model, owns planning, ambiguity and review; the cheaper sidekick explores, implements, runs tests and reports back from the lead's briefs. They exchange only briefs, results and feedback, so both benefit from caching. Cognition found pricier but more token-efficient models can lower total cost, and tunes brief detail and sidekick pushback per model pairing.

Benchmarks

Artificial Analysis data shows Fusion with Opus 5.5 Max + GPT-6 Luna High at 68.1: DeepSWE v1.1 72.6, SWE-Atlas-QnA 69.6, Terminal-Bench v4 62.1. Muse Code with Muse Wonder Full is second at 68.0, effectively tied; Antigravity CLI with Gemini 4 Argon scores 63.8 and Codex with GPT-6.1 Sol (xhigh) 62.9. The 31% lower cost per task vs Claude Code with Opus 5.5 is Cognition's claim.

Getting started

Install Devin CLI with curl -fsSL https://cli.devin.ai/install.sh | bash, pick Fusion, and choose a lead and sidekick; Cognition recommends Opus 5.5 + GPT-6 Luna. It suits long, multi-file repo tasks; compare cost per task on your own codebase rather than price per token.