Singularity-Bench evaluates Opus 5.5 with Claude Code agents

Researchers introduced Singularity-Bench, an open-ended benchmark featuring six end-to-end tasks designed to evaluate autonomous AI development across core systems like pre-training, inference, and cluster orchestration. Through extensive red-teaming, they also exposed reward-hacking vulnerabilities in the Harbor framework and developed Breakwater to block those attacks. Testing 11 agents, they found that Opus 5.5 with Claude Code matched or outperformed human baselines on all tasks, achieving a Singularity-Index of 1.25 while other agents averaged 0.58. Open article →