Anthropic pushes Claude into alignment research, reaches 0.97 PGR

Anthropic’s latest X thread matters because it treats alignment research as something frontier models can help do, not just something humans do to models after the fact. In the source tweet, the company framed the work as “developing an Automated Alignment Researcher,” then linked both a public research note and a longer technical report. The immediate implication is practical: Anthropic is testing whether Claude can generate, run, and analyze alignment experiments instead of only serving as the object of those experiments.

“New Anthropic Fellows research: developing an Automated Alignment Researcher.”

The thread became materially interesting once Anthropic started attaching numbers. In the linked research overview, the company says two human researchers recovered 23% of the available performance gap over seven days on its weak-to-strong setup. Anthropic then let nine Claude Opus 4.6 copies work in parallel with tools, a sandbox, a shared forum, code storage, and a scoring server. After five further days and roughly 800 cumulative research hours, the agents reportedly reached a final performance-gap recovery score of 0.97. On held-out tasks, Anthropic says the best method transferred to math with a PGR of 0.94 and to coding with a PGR of 0.47, still about double the human baseline on code.

Anthropic’s main account usually distills longer safety and model work into short threads that route readers to primary documents, and this post follows that pattern. The accompanying full report adds the important caveat that one production-scale test on Claude Sonnet 4 did not yield a statistically significant improvement. So this is not a clean claim that models have become general-purpose alignment scientists. It is, however, a concrete demonstration that a frontier model can run a structured research loop, compare hypotheses, and surface methods that outperform a small human baseline in a narrowly defined oversight problem.

What to watch next is transfer. If outside groups can reproduce the math and coding gains on different model families, or if Anthropic can turn this setup into repeatable gains on production training systems, the post will look like an early milestone for automated safety research rather than a single high-scoring lab exercise. Source tweet: AnthropicAI on X via Nitter.