Whether AI agents can build and strengthen AI R&D infrastructure on their own is becoming an important measure of frontier models’ autonomous capabilities. Researchers from Carnegie Mellon University and other institutions recently introduced the Singularity-Bench benchmark, which includes six tasks related to AI R&D to assess agents’ ability to autonomously build AI R&D capabilities. In this test, Claude Opus 5.5 paired with Claude Code delivered impressive results.

Latency Sharply Reduced, Scheduling Optimization Boosts Hardware Output

At the model inference performance level, this combination demonstrated quite outstanding engineering optimization results. Under the same hardware conditions, Claude’s self-developed inference engine sharply reduced P90 first-token latency from about 12 seconds with the mature inference engine SGLang to 0.48 seconds.

In addition to a leap in first-token response speed, the inference engine further improved overall processing efficiency and GPU output performance through optimized scheduling and other methods.

Five Wins and One Tie in Six Tasks, Surpassing the Human Baseline; Self-Improvement Remains in Early Exploration

In the six AI R&D tasks set by Singularity-Bench, the combination of Claude Opus 5.5 and Claude Code ultimately achieved a record of five wins and one tie, becoming the only entrant whose total score exceeded the human baseline.

However, the researchers also pointed out the limitations of the current test: at this stage, the scope of the test tasks is still defined by humans, and true AI recursive self-improvement still awaits further exploration and validation.