Can AI actually create new science, or just help us do ours faster?

I've been sitting with this question for a few days now.
Most "AI for science" talk is still about faster lit reviews, cleaner code, tidier plots. Useful, but not new. What Anthropic just showed with Mythos-class models feels like something else.
In internal tests, Mythos 5 sped up drug design roughly 10x. Nine of fourteen protein targets it worked on produced strong candidates. And in blind comparisons, researchers preferred its molecular biology hypotheses over Opus-class models about 80% of the time.
But the genomics result is the one that actually stopped me mid-scroll. Mythos 5 pulled together single-cell data across 138 species, designed its own modeling approach, trained something 100x smaller than the comparison system — and that smaller model beat a result published in Science. Not "helped write it." Ran the whole thing: hypothesis, method, execution, evaluation.
As someone who's spent years training models, writing papers, and knowing exactly how much grinding sits behind every "novel" result — that one made me pause.
And notice: access is still tightly locked down. Project Glasswing, a small biology trusted-access program coming soon. That's not a company hyping a headline. That's a company that knows it's holding something it isn't fully sure how to hand out yet.
So is this "new science"? Not a new field, not new axioms. Not yet. But autonomously forming hypotheses and beating specialists at their own game — that's not incremental anymore.
I don't have a clean answer. I just know the line between "tool" and "collaborator" moved this week, and I'm not sure my mental model of research has caught up.
Where do you draw that line?