Current frontier models are not good enough because they still suffer from major hallucinations, sycophancy, and context drift. So there has to be at least (and

Current frontier models are not good enough because they still suffer from major hallucinations, sycophancy, and context drift. So there has to be at least (and I have no reason to believe it will be the last, GPT-5 demonstrates that the transformer architectures are hitting diminishing returns) one more training cycle.