Anthropic had Claude align other AI models in 48 hours on one GPU

Anthropic handed Claude a single GPU, 48 hours and a hard problem: make other AI models safer, on its own.

What the research showed

In Fellows research published on August 28, 2026, Claude autonomously researched alignment methods, then trained and tested small models to apply them, with no human in the loop. Anthropic reported that it worked surprisingly well. In a second test, the smaller Sonnet 5 post-trained an early checkpoint of the more capable Opus 4.8 and pushed its safety scores close to those of the fully aligned production model. Anthropic released the automated alignment setup so other researchers can build on it.

Why it matters

This is one of the first concrete demonstrations of AI improving the safety of AI, including a weaker model steering a stronger one. If automated alignment scales, it could expand oversight faster than human researchers can. Anthropic was careful about the limits: the approach only works when researchers measure the right things, because subtle or rare failures may have no benchmark at all. The open question is whether automated evaluations can catch the failures that no existing test is designed to surface.