Found first: a primary source the press has not covered yet.
Researchers at the University of Science and Technology of China, StepFun, Peking University, HKUST, Yale, and the University of Pennsylvania have published Φ-Bench, a benchmark for evaluating whether frontier LLMs can engineer the infrastructure that runs them. Across 85 tasks drawn from real research repositories and top systems conference papers, Claude Opus 5 leads all eight tested models with an overall score of 36.53%.
What the source says
The benchmark spans nine infrastructure categories, from training and inference serving to hardware and edge deployment, built from 2,260 papers and 1,852 code artifacts from top-tier systems conferences between 2023 and 2026. Tasks fall into three formats: kernel function completion (KFC), long-horizon implementation (LHI), and end-to-end optimization (E2EO). Claude Opus 5 scored 37.16% on KFC, 21.60% on LHI, and 62.94% on E2EO. On Hardware and Edge tasks specifically, Claude Opus 5 scored 3.90%. The next best overall scores were Kimi K3 at 28.12%, Qwen3.8 Max at 27.73%, GPT-5.6 Sol at 24.51%, and GLM 5.2 at 21.92%.
Why it matters
The 3.90% score on Hardware and Edge tasks puts a near-zero figure on one of the most specific versions of the self-improving AI premise: that frontier models will eventually handle the silicon-level work of deploying and tuning the systems that run them. Long-horizon implementation, the format most analogous to open-ended autonomous development, sits at 21.60% for the best model. The E2EO score of 62.94% looks stronger, but that task type provides a constrained optimization target rather than an open-ended specification. The benchmark draws from actual frontier research repositories, which makes the gap harder to attribute to synthetic or narrowly constructed test conditions.