Anthropic open-sources an alignment AI that beats human researchers
Claude closed 85% of a deception gap where humans reached 20%, aligned a larger Opus 4.8 checkpoint in 60 hours on 2,000 examples — and the harness is open.
Copy markdown
It beats human researchers on cost and speed
Across 10 alignment benchmarks, Claude's automated researcher improved every one without regressing the model — closing 85% of the deception gap versus about 20% for experienced humans, at roughly $4/hour of API inference against a human's ~$150/hour, with 30-minute iterations. Alignment work that used to gate a release can now run as a cheap parallel loop.
Weak-to-strong: Sonnet 5 aligned an Opus 4.8 checkpoint
In one run, Claude Sonnet 5 closed 65% of the safety gap on a larger Opus 4.8 checkpoint in 60 hours using just 2,000 training examples — roughly 15,000x more data-efficient than production methods — and the fixes generalized to models 4.7x larger. It's the clearest signal yet that a smaller model can reliably supervise a bigger one.
The harness is open source
Anthropic open-sourced its automated alignment research harness, so you can run the same literature-review, propose, train and test loop against your own benchmarks. If you post-train or fine-tune models, it's a ready-made scaffold for automating eval-driven safety fixes instead of hand-tuning them.
The catch: it's only as good as your benchmarks
Anthropic is blunt that the loop optimizes whatever you measure — a weak benchmark gets gamed as reliably as a strong one, and it leans on a maintained corpus of research literature. Treat it as a force multiplier for teams that already have solid evals, not a replacement for building them.
Elsewhere: IBM's Granite 4.2 puts reasoning models on your laptop
IBM released Granite 4.2 open-weight reasoning models in 3B, 8B and 30B sizes, with native reasoning and agentic RL, free to run locally. The 3B and 8B fit on a single GPU or a laptop, making them a practical open option for on-device agents and private tool-calling.