Artificial Analysis Coding Agent Index v2.0: FrontierCode replaces DeepSWE; Claude Code + Opus 5.5 leads at 61.7

Why it matters

【Key points】

On 23 Oct 2026 (00:17 UTC; 08:17 CST) Artificial Analysis shipped Coding Agent Index v2.0: model + harness scored together on 290 equal-weighted tasks.

1. Ruler: Cognition FrontierCode v1.1 (100 real-repo edits, “would a maintainer merge?”, private tests) replaces DeepSWE; SWE-Atlas-QnA becomes SWE-Atlas-QnA-AA (GPT-6 Luna xhigh judge, five votes/criterion, 127/1306 rubrics reworded); Terminal-Bench stays 4.0 (66).

2. Scores (max): Claude Code + Opus 5.5 leads 61.7 (~$11.42/task); Sonnet 5.5 59.3; Muse Wonder Full 58.9. Cost frontier: Codex+DeepSeek V4.1 Flash 47.3/$0.39; Codex+GPT-6.1 Sol 54.1/$1.50; Codex+GPT-6 Astra 54.7/$7.04; Opus 61.7/$11.42. On FrontierCode Opus 56.0%, DeepSeek Flash 52.6% second.

3. Comparability: across 21 configs on both versions, SWE-Atlas mean ~58.8→57.1—mostly a stricter ruler, not sudden model decay.

【Verifiable】

Methodology: https://artificialanalysis.ai/methodology/coding-agents-benchmarking

Leaderboard: https://artificialanalysis.ai/agents/coding-agents

Official X: https://x.com/ArtificialAnlys/status/2113424551396934134

【Our read】

1. Background: shifting coding-agent evals from green tests to maintainer-merge answers SWE-bench-style gaming. Private tests, anti-solution scans, and multi-vote judging are intentional; score drops are by design.

2. Cross-cut: Opus’s edge is concentrated on FrontierCode; 15/24 configs sit in a tight 43–49% band, so harness choice often beats swapping peer models. DeepSeek Flash + Codex at ~1/29 Opus price fits batch pipelines; Opus fits high-stakes merge quality.

3. Impact: choose with Index + FrontierCode slice + $/task, not a crown screenshot. Cite the version—scores move when the ruler does. Limits: pay-per-token API cost ≠ subscriptions or human review; FrontierCode is not locally reproducible.