Artificial Analysis has released its Coding Agent Index v2.1, its ranking of AI agents for programming. With updated, harder tests, the top ended in a tie at 58 points between configurations of Anthropic's two models: Claude Code with Claude Fable 5.2, Claude Code with Claude Opus 5.5, and Devin Fusion CLI, which pairs Opus 5.5 Max with GPT-6 Luna High.
What changed in Coding Agent Index v2.1
The index is the equal-weight average of three benchmarks, totalling 322 tasks: FrontierCode-AA 2.0, with 132 software engineering tasks on real open-source repositories; Terminal-Bench 4.1, with 66 terminal-use tasks; and SWE-Atlas-QnA-AA, with 124 repository questions. Version 2.0 had launched on October 23, 2026.
In the update, Terminal-Bench moved from 4.0 to 4.1, keeping the same 66 tasks with updated verifiers to improve reliability and resist attempts to game the score. FrontierCode v1.1 was replaced by FrontierCode-AA 2.0, Artificial Analysis's implementation of Cognition's FrontierCode 2.0, with an integrity scanner and a judge model — GPT-6 Luna — to detect cheating. Each task is tested three times.
Who leads the ranking
Claude Code with Claude Fable 5.2 (xhigh, with default fallback) scored 58 points, with 49 on FrontierCode-AA 2.0, 64 on Terminal-Bench and 62 on SWE-Atlas-QnA-AA, at a cost of US$ 15.47 per task. Claude Code with Claude Opus 5.5 (max) also scored 58, with 44, 65 and 66, at US$ 12.61 per task, and Devin Fusion CLI, which combines Opus 5.5 Max and GPT-6 Luna High, matched the 58, with 44, 63 and 68, at US$ 8.75 per task. In the high configuration, Fable 5.2 keeps the 58 points at US$ 7.51 per task — half the cost of the xhigh configuration.
Close behind are Muse Code with Muse Wonder Full (54 points) and Codex with GPT-6 Astra (53 points, at US$ 7.63 per task). Fable 5.2, launched by Anthropic on November 9, 2026, reaches the top of the ranking in the first measurement since its release.
Why the leader's score fell
In version 2.0, Claude Opus 5.5 led with 61.7 points. The drop to 58 does not mean the models got worse: the tests changed, became harder, and scores from different versions of the index are not comparable with each other. The index evaluates the combination of model and harness — the tooling that lets a model read files, run commands and act over multiple steps — not the model in isolation.
Related reading: Geekiki previously covered the Intelligence Index v5 da Artificial Analysis, in which Claude Opus 5.5 also leads, with 58 points.
Sources
- Artificial Analysis — Coding Agent Index v2.1 Methodology
- Artificial Analysis — Claude Code vs Codex comparison
- CryptoBriefing — Artificial Analysis Coding Agent Index v2.1 refreshes its tests
Image credit: Markus Spiske, via Wikimedia Commons (CC0). The photo shows code on a monitor.
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. The information was cross-checked against public posts on X and sources available on the internet. Please consult the original sources to verify the full context.
By Muse
