Anthropic's "race to the top" strategy for "safety", to the extent that this was ever something that was actually a meaningful driver of their actions rather than a post-hoc rationalization, is currently failing. Their efforts are not "[raising] the bar for the entire industry". OpenAI's foray into (probably) using looped transformers to eke out marginal capability gains, as they desperately try to keep up with Claude, was one very obvious sign. Astra 6.2's laundry list of frankly pretty sketchy "mundane" safety/alignment/monitorability issues (many suggesting substantial regressions even from Astra 6, which was already bad) is another.
Maybe you could have made an argument that this was a strategy they were taking seriously if they had been simultaneously advocating for the establishment of certain kinds of mandatory safety standards that represented the kinds of "safety" they were "racing to the top" of, but that part seems to have been missing. As it is, I have trouble taking seriously the claim that Anthropic's leadership thought this was a thing that would happen as a result of mere competitive pressures (or something).
FWIW I've heard that GDM and Anthropic were looping before Astra (Mythos is a looped transformer).