The biggest jump was Opus 4.6. Since then they have gradually gotten better at finding issues in your reasoning, not hallucinating, and being rigorous with the

The biggest jump was Opus 4.6. Since then they have gradually gotten better at finding issues in your reasoning, not hallucinating, and being rigorous with the code, but much much worse at explaining things and generally just talking in a way that a human can understand. All the models I've tried seem to be suffering from the same fate so it must be something going on with the training meta right now.