Devin Fusion matches top coding agents at lower cost per task

A lead-plus-sidekick configuration scored 68.1 versus 66.0 for plain Claude Code on Opus 5.5, at $9.01 versus $13.04 per task. Splitting work across a frontier and a cheaper model can cut agent cost without losing accuracy.

• Devin Fusion runs a frontier lead model and a cheaper sidekick in parallel, each with its own context and tools, while the lead keeps planning, delegation and final review. Artificial Analysis tested Claude Opus 5.5 (max) leading GPT-6 Luna (high).
• The pair scored 68.1 on the Coding Agent Index, closely matched with Claude Code on Sonnet 5.5 (max) at 68.4 and Muse Code at 68.0, and averaged 55 minutes per task against 65 for Claude Code on Opus 5.5 (max).
• The gain is uneven: Fusion beat Claude Code on Opus 5.5 on DeepSWE v1.1 (72.6% vs 68.4%) and SWE-Atlas-QnA (69.6% vs 66.4%) but trailed on Terminal-Bench 4.0 (62.1% vs 63.1%). Reported costs are combined pay-per-token API costs, not subscriptions.

• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
• context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
• token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

• postAA-PersonalityBench ranks frontier models by sycophancy
• postArtificial Analysis rebuilds its medical long-context leaderboard
• postDeepSeek V4.1 Flash 1017 jumps six points and ties for #2 open weights
• postMistral's Voxtral 3 speech models benchmarked: top-3 STT, mid-pack TTS

• postAA benchmarks Devin Fusion, first multi-model coding agent on its indexArtificialAnlys
• postVals AI Finds Devin Fusion on the Pareto Frontier for Code MigrationVals AI
• postCoding Agent Index: Sonnet 5.5 leads on score, GPT-6.1 Sol wins on costArtificialAnlys

Checking sign-in…

Loading comments…