Artificial Analysis 编程智能体指数新增安全拒答与回退模型报告
Safety refusal reporting in the Artificial Analysis Coding Agent Index now shows when refusals occur and which models are used as fallbacks
Artificial Analysis 的 Coding Agent Index 新增安全拒答报告，可查看拒答发生在任务提示词阶段还是任务中途，以及拒答后智能体切换到了哪个回退模型。Claude Code 搭配 Sonnet 5.5（max）目前排名第一，拒答率 4.5%，约为 Claude Code 搭配 Opus 5.5（max）8.9% 的一半；Sonnet 5.5 约 94% 的拒答发生在首轮之后，拒答后几乎总是切换到 Opus 4.8。
正文
当前显示原文，中文全文译文尚未就绪。
Safety refusal reporting in the Artificial Analysis Coding Agent Index now shows when refusals occur and which models are used as fallbacks
We've added two ways to explore the results:
➤ Refusal timing: see whether a refusal occurred from the task prompt alone or later in the task, after the agent had already started working
➤ Fallback model: see which model the agent switched to after a refusal, alongside attempts that were blocked
Claude Code with Sonnet 5.5 (max) currently ranks first in the Index. Its safety refusal rate is 4.5%, roughly half the 8.9% recorded for Claude Code with Opus 5.5 (max).
Around 94% of Sonnet 5.5's refusals occurred after the first turn. Following a refusal, the agent almost always switched to Opus 4.8.
Observed fallback patterns vary across model and agent configurations. With Fable 5.1, Claude Code fell back predominantly to Opus 4.8, while Opus 5 accounted for a much larger share of Devin Fusion's fallbacks in the Index.
Both views are available for the overall Index and each benchmark. Safety and fallback behavior is provider-configured and may change over time; these rates reflect behavior recorded at the time of benchmarking.