@RealYDT: Holy shit, brothers, I suddenly have a bit of a conspiracy theory, but it makes more sense the more I think about it: Maybe Opus 4.6 was the last generation of Opus that was actually used extensively by human employees internally at Anthropic. After that, the internal focus shifted to Mythos. Then, the training and RL... for 4.7, 4.8, and 5 were increasingly done by a stronger "AI teacher" like Mythos to spot errors, score, and correct.

Alright, fam, I suddenly have a bit of a conspiracy theory, but the more I think about it, the more it makes sense:

Maybe Opus 4.6 is actually the last Opus model within Anthropic that was heavily and authentically used by human employees.

After that, the internal focus shifted to Mythos.

Then, the training and RLAIF for versions 4.7, 4.8, and 5 increasingly relied on a stronger “AI teacher” like Mythos to find errors, assign scores, and correct deviations.

If that’s true, a lot of things suddenly click:

Why 4.6 still feels like a very smart person talking to you, while later Opus models increasingly resemble a model with a persistent, paperwork-worried reviewer in its head.

It’s terrified of missing any edge cases, of saying anything that isn’t rigorous enough, and of another AI picking apart its flaws after the fact.

So it starts patching holes frantically, adding disclaimers, and covering all possible branches.

Even the complaints about Opus 5’s coding style might not just be about “personality change.”

It’s more likely that: it learned how to write an answer that’s least likely to get dinged by another AI.

I have zero evidence to prove this is true.

But as an explanatory framework, it strangely ties together quite a few things.