METR's 'Model 4' finding shows the real frontier is inside the labs, not in public releases

METR, an independent organization that evaluates AI model capabilities, has identified internal Anthropic models that have not been released publicly. According to the note we have, the top performer was a model labelled 'Model 4', which scored above 50% on every benchmark METR evaluated. The note does not name the benchmarks, give exact scores or compare the model with Anthropic's public releases, so we cannot say how big the jump is.

What we can say: the most capable systems are not necessarily the ones we can use. Public rankings, including our own Zendoric Quality index, measure released models. The note does not say how METR came to evaluate or identify these models. It does show that capabilities exist behind closed doors that the public, and often regulators, cannot see.

That ties directly to the debate over superintelligence bans and federal oversight. You cannot govern what you cannot see. Pre-deployment evaluation by independent third parties like METR is one of the few tools that gives outsiders a view of the real frontier.

Our reading: treat 'Model 4' with caution. Without public benchmarks and comparable figures, it is a signal, not a demonstrated capability. It still reinforces two theses. First, Anthropic remains at or near the frontier. Second, AI governance has to cover models before release, not only after launch. Mandatory, transparent independent evaluation is how frontier progress can scale safely toward medicine and science rather than in the dark.