Here, we measure success by the fraction of the “performance gap” we can close b

Anthropic's Automated Alignment Researchers (AAR) program has produced a striking empirical result: Claude Opus 4.6, equipped with additional tools, closed 97% of the "performance gap recovery" (PGR) metric in a weak-to-strong supervision experiment, compared to the 23% human researchers achieved after seven days of work. The experiment defined success by measuring how much of the gap between a weak "teacher" model — Qwen 1.5-0.5B-Chat — and a stronger "student" model — Qwen 3-4B-Base — could be closed through automated generalization methods. Human researchers established the baseline PGR of 0.23, while the Claude-powered AARs reached a PGR of 0.97 over five additional days, accumulating 800 research hours at a total cost of approximately $18,000, or roughly $22 per AAR-hour. The result represents one of the most quantitatively concrete demonstrations to date of AI systems materially accelerating alignment research rather than merely assisting with it. The significance of this finding extends well beyond the raw numbers. Weak-to-strong supervision is considered one of the central unsolved problems in AI safety: as models grow more capable, the humans and systems tasked with supervising them become increasingly unable to verify whether outputs are actually correct or aligned. The recursive structure of the AAR approach — using a capable AI model to research how to better supervise AI models — is either an elegant solution to this bootstrapping problem or a demonstration of its limits, depending on how the research holds up under scrutiny. Observers in the broader AI community have noted that the critical unresolved question is whether a model can reliably identify flaws in its own training regime, a task that requires the supervisor to reason about failure modes it may have been trained not to surface. The cost-efficiency dimension adds another layer of significance. At $22 per research hour, the AAR approach suggests that certain categories of alignment research can be industrialized and scaled in ways that human researcher pipelines cannot match. This matters because alignment research has historically been constrained not only by conceptual difficulty but by the sheer labor intensity of empirical iteration. If AARs can reliably compress the timeline between hypothesis generation and empirical validation, the field gains a compounding advantage — each generation of alignment insight potentially feeding faster into the next. However, researchers and commentators have noted that the gap between experimental performance and production transfer remains a critical open question, with results that close benchmark gaps not automatically translating into robust real-world supervision reliability. This development fits within a broader pattern of AI labs treating AI systems themselves as research accelerants, a trajectory sometimes called "recursive improvement" or "AI for science." Anthropic's AAR work is among the most explicit instantiations of this approach applied specifically to safety research, which carries both promise and irony: the same capability advances driving the need for better alignment tools are being recruited to produce those tools faster. The community response has been largely attentive to this tension, with observers noting that measurable supervision quality — not just raw model capability — is the practical unlock that determines whether the approach scales. Whether the PGR metric proves robust enough to serve as a reliable proxy for real-world alignment progress will likely be a central question as this line of research matures into 2026 and beyond.