At 08:17 Beijing time on October 23, 2026, Artificial Analysis announced Coding Agent Index v2.0. The headline change is methodological: Cognition's FrontierCode v1.1 replaces DeepSWE, SWE-Atlas-QnA is revised into SWE-Atlas-QnA-AA, and Terminal-Bench 4.0 is unchanged. The Index still averages three benches with equal weight across about 290 tasks (100 + 66 + 124). DigIntoAI has not re-run the index.
On the live Coding Agent Index page checked for this article, Claude Code with Claude Opus 5.5 (max) leads at 61.7, with an average pay-per-token API cost of about $11.42 per task. Devin Fusion CLI (Claude Opus 5.5 Max + GPT-6 Luna High) sits close behind at about 61.5 on the same board. Those figures are Index scores under the new component set, not a claim that older DigIntoAI write-ups on the prior board still rank the same way—see Best Coding Agents Right Now, Ranked by Artificial Analysis (October 2026) and Devin Fusion Scores 68.2 on Artificial Analysis Coding Agent Index, Level With Claude Code and Muse Code for prior context, noting those pieces predate v2.0.
FrontierCode is where Opus 5.5 pulls ahead on the software-engineering slice: Artificial Analysis reports 56.0% for Opus 5.5 on FrontierCode v1.1, ahead of Codex with DeepSeek V4.1 Flash (max) at 52.6%. Because FrontierCode replaces DeepSWE and grades maintainer-merge style changes on a private task set, Index total scores should not be hard-compared across Index versions—the new suite is a different yardstick, not a quiet model regression.
Separately, Artificial Analysis's announcement says that across 21 configurations scored on both Index versions, the SWE-Atlas mean falls from 58.8% to 57.1% under the stricter SWE-Atlas-QnA-AA grading (GPT-6 Luna xhigh judge, five votes per criterion). That is a per-bench note from the announcement, not a claim about the composite Index mean.
Costs cited here are pay-per-token API costs, not subscription prices. DigIntoAI has not reproduced the leaderboard runs.