It's not a skill issue.
I have a hypothesis:
There's been a clearly observable design change in the way the model is built.
Meta’s mistake with Llama 4 was easier to see because the failure was basically sitting there with a little flag on it. You can isolate it. You can test it. Visual reasoning only travels so far through the model because the design appears to prevent modal bleed from crossing deeply into general reasoning. And the projector they used for it does not seem strong enough to compensate.
My concern is that Anthropic may have made a related tradeoff in Opus 4.7. Not necessarily the same implementation, but the same kind of mistake: treating modal bleed as contamination instead of as part of the model’s reasoning texture.
Thing is, the cross modal reasoning change meaningfully changes the behavior of the model. Every claude model since Sonnet 3.5 has done it.
The design changes lead to shallower, less rich reasoning on all but a few very specific tasks, and it hurts the model's ability to cross reason.
And you can test it by checking on the way cross reasoning against code works. Comparing Opus 4.1 to Opus 4.7, you'll notice a huge style difference if you tell it to code against non-code related priors. It's night and day.
Failure modes are easiest to spot in weird fizzbuzz prompts.
If you get the pattern matched textbook default, the model has failed to cross-reason about the problem.
Gemini models still bleed, though.
And there are a few open weights models like Kimi K2x that seem to do it the same way Claude did.