i get the sense that anthropic has already used its Model 3 to reach Astra-level performance with opus 5.5, something its Model 2 couldn't manage with fable 5.1
and openai held back its own Model 3 equivalent which suggests it may actually be a generation ahead internally
these AECI scores suggest to me that Model 4 is only as capable as aug? Bel?, despite being trained 1-2 months later