Devin Fusion tops Coding Agent Index with score of 68

On October 21, 2026, Artificial Analysis posted on X that Cognition's Devin Fusion, running Claude Opus 5.5 (max) as its lead model with GPT-6 Luna (high) as a lower-cost sidekick, scored 68 on the Coding Agent Index. That was 2 points above Claude Code with Opus 5.5 (max), at 31% lower cost per task.

Two days later, on October 23, Artificial Analysis released Coding Agent Index v2.0. The new version replaces DeepSWE with Cognition's FrontierCode v1.1 and revises grading for SWE-Atlas-QnA. On the leaderboard, Claude Opus 5.5 in Claude Code leads at 61.7. According to the organization, FrontierCode v1.1 contains 100 software engineering tasks drawn from real open-source repositories, uses a private test set, and grades results on whether a maintainer would merge the change. Claude Code with Claude Opus 5.5 (max) scores 56.0% on that benchmark.

In the same announcement, Artificial Analysis said scores under v2.0 are lower than under v1.5 for almost every configuration measured on both. Across the 21 configurations evaluated on both versions, the mean Index score falls from 54.8 to 47.4, with 20 of the 21 moving down. The organization attributes the shift to the replacement of the old benchmark with FrontierCode v1.1. The v2.0 index uses FrontierCode v1.1 together with the SWE-Atlas-QnA-AA dataset and judge, and the leaderboard lets users view each configuration's performance by benchmark score, cost per task, and token usage, alongside methodology and version history. The public record currently stops at the launch of this leaderboard.