Artificial Analysis launches Coding Agent Index v2.0: FrontierCode from Cognition replaces DeepSWE, and Claude Code with Opus 5.5 leads at 61.7

Background and the problem

A coding agent's results depend on the harness (Claude Code, Codex, Muse Code and others) as well as the model, so Artificial Analysis scores each model-plus-harness pair as its own row. In v1.5, top configurations were already above 65% on DeepSWE v1.1, which made it less useful for telling them apart. Version 2.0 swaps in a harder software engineering benchmark that is closer to real merge standards.

How the index works

Version 2.0 averages three equally weighted benchmarks over 290 tasks: FrontierCode v1.1 from Cognition (100 tasks, private test set), Terminal-Bench 4.0 (66 terminal tasks, unchanged) and SWE-Atlas-QnA-AA (Scale AI's 124 repository Q&A tasks). FrontierCode grades whether a maintainer would merge the change: any failed blocking check scores zero, and passing all of them earns 70% to 100% depending on non-blocking checks. Agents have internet access, but a trajectory scan zeroes any attempt that fetched the upstream pull request, merge commit or a mirror. SWE-Atlas-QnA-AA rewords 127 rubric criteria and uses GPT-6 Luna (xhigh) as judge with five votes per criterion. Each task is run three times and averaged.

Benchmark results

All figures are from Artificial Analysis. Overall: Claude Code with Opus 5.5 (max) 61.7, Sonnet 5.5 (max) 59.3, Muse Wonder Full (max) 58.9, Fable 5.1 57.6. FrontierCode alone: Opus 5.5 56.0%, Codex with DeepSeek V4.1 Flash 52.6%, GPT-6.1 Sol 49.4%, GPT-6 Astra 49.3%. The index got harder: 21 configurations fell from a mean of 54.8 to 47.4, and 20 of them moved down. On the cost frontier, DeepSeek V4.1 Flash scores 47.3 at $0.39 per task, while Opus 5.5 scores 61.7 at $11.42.

What developers can do

Don't pick on the headline score alone. For high-volume or budget work, try Codex with DeepSeek V4.1 Flash or GPT-6.1 Sol first, and move to Opus 5.5 when merge quality matters most. Scores on v2.0 are not comparable with v1.5, so check the version before comparing older articles. Per-benchmark scores, cost per task and token usage are on the Artificial Analysis coding agents page, and the SWE-Atlas-QnA-AA dataset can be downloaded from Hugging Face for your own runs.