Devin Fusion matches top coding agents at lower cost per task

Artificial Analysis, which runs its own benchmarks of coding agents, tested Cognition's Devin Fusion command-line agent configured with Claude Opus 5.5 at max effort as the lead model and GPT-6 Luna at high effort as a cheaper 'sidekick'. It scored 68.1 on the firm's Coding Agent Index. Claude Code running Opus 5.5 by itself scored 66.0. According to Artificial Analysis, the Fusion run cost $9.01 per task against $13.04 and averaged 55 minutes per task against 65. That puts it in a near tie at the top with Claude Code using Sonnet 5.5 at 68.4 and Muse Code using Muse Wonder Full at 68.0.

The setup splits one coding job across two models from different labs, Anthropic for the lead and OpenAI for the sidekick. As Artificial Analysis describes it, the two run in parallel, each with its own context and tools, and the lead keeps planning, delegation and final review. The reported score and cost cover both models working together, including their combined API bill. The post does not say how much of the work the sidekick handled. The comparison also changes the agent software itself, Cognition's tool versus Claude Code, so the result shows what this whole configuration achieved, not how much of the saving comes from the second model.

The index is a simple average of three benchmarks, and each benchmark score is the average pass rate over three attempts per task, according to the Coding Agent Index page. DeepSWE v1.1 covers software engineering tasks, Terminal-Bench 4.0 covers multi-step work in a shell, and SWE-Atlas-QnA covers technical questions about a codebase. Fusion beat solo Opus 5.5 on DeepSWE, 72.6% to 68.4%, and on SWE-Atlas-QnA, 69.6% to 66.4%. It came in slightly lower on Terminal-Bench, 62.1% to 63.1%. Costs are pay-per-token API prices, and the page notes that many people use these tools through subscription plans instead, so the dollar gap shows relative token spend, not what a subscriber pays.

This is not Artificial Analysis's first Devin Fusion run. In September it benchmarked the launch configurations, which paired a frontier lead with Cognition's own SWE-2 model as the sidekick and scored 62 with Claude Fable 5.1 as the lead. The $13.04 comparison figure comes from the firm's earlier Opus 5.5 test, which found that Opus 5.5 topped the index at the time but used substantially more tokens per task than its predecessor. In the new run, Opus 5.5 working inside Fusion with a cheaper sidekick came out slightly ahead on score and well below on cost.