CMU 论文：Opus 5.5 + Claude Code 搭建的系统在 Singularity-Bench 上超越 SGLang
CMU 这次不考答题了，直接让智能体搭整套生产系统跟 SGLang、NVIDIA Dynamo 硬碰硬。Opus 5.5 做出的推理引擎每 GPU 小时赚 11.52 美元，比 SGLang 的 8.18 还高。做智能体评测和基础设施的工程师都该看看评分方式。
卡内基梅隆大学发布新论文 Singularity-Bench，给智能体布置 6 项端到端工程任务，包括预训练一个模型、搭建推理服务引擎和集群编排器，每项结果都和 SGLang、NVIDIA Dynamo 等生产级系统在同一批 GPU 上对比打分。
测试中 Opus 5.5 加 Claude Code 实际搭出了预训练管线、推理引擎和集群编排器三项完整系统，其推理服务引擎每 GPU 小时产出 11.52 美元，对比 SGLang 的 8.18 美元高出约 40%。
在总指标 Singularity-Index 上，Opus 5.5 拿到 1.25（1 等于人类基线），其余 10 个智能体平均只有 0.58。这篇基准把智能体评测从问答和代码片段推进到完整生产系统层面，对评估智能体的真实工程能力有参考价值。
New Carnegie Mellon paper finds that Opus 5.5 + Claude Code built a pre-training pipeline, an inference engine, and a cluster orchestrator that matched or beat production tools like SGLang and NVIDIA Dynamo on the same GPUs.
Singularity-Bench gives agents 6 end-to-end jobs, from pre-training a model to building a serving engine. Each result is scored against a production system, such as SGLang or Claude Code, on the same GPUs.
Opus 5.5 scored a Singularity-Index of 1.25, where 1 equals the human baselines, while the other 10 agents averaged 0.58. Its serving engine earned 11.52 USD per GPU-hour against 8.18 for SGLang.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Opus 5.5, Claude Code, SGLang, Singularity-Bench.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
