Artificial Analysis announced Coding Agent Index v2.1 on November 14, with Claude Fable 5.2 and Claude Opus 5.5 configurations sharing the lead at 58. In Claude Code, Fable reaches that score at high and xhigh; Opus reaches it at max. Source
For Fable in Claude Code, max costs more than high or xhigh but scores lower because more attempts hit task time limits, according to the evaluator. That’s an expensive way to finish behind yourself. Source Source
Cognition’s Devin Fusion CLI also reaches 58 with a multi-model configuration pairing Claude Opus 5.5 Max and GPT-6 Luna High. These configurations share the lead at the announcement’s whole-number precision. Source Source
New tasks, lower scores
The refresh changes two of the index’s three components, but in different ways:
- FrontierCode-AA 2.0 replaces FrontierCode v1.1. Artificial Analysis describes its implementation of Cognition’s benchmark as 132 real-world coding tasks, graded on whether a repository maintainer would merge the agent’s change. Source
- Terminal-Bench 4.1 replaces version 4.0. It retains the same 66 tasks but updates their verifiers, with the stated aim of improving reliability and resistance to reward hacking. Source
- SWE-Atlas-QnA-AA stays in the index. Source
Artificial Analysis reports that overall index scores are five points lower on average, driven by the FrontierCode upgrade. Across the 23 configurations tested on both FrontierCode versions, every configuration scored lower, with the mean falling from 45% to 29%. Those are different task sets, not evidence that the models suddenly forgot how to code. Source Source
Terminal-Bench’s verifier refresh produced smaller changes: across the 23 configurations evaluated on both versions, every score moved by less than three points. Source
Fable max spends more to score less
Artificial Analysis reports the following costs and index scores for Claude Fable 5.2 in Claude Code:
Comments