Prompt clarity

Anthropic just released Opus 4.7, and it's had mixed reviews. It seems they might have rushed this model to try to get ahead of opening eyes release of Codex for everything.

People also suspect OpenAI is about to release their next big model, spud. And if Google doesn't have a big update soon, they certainly will next month at Google Io.

All three of these model makers are trying to be the model that you use for work, and the race has never been more tight. Let's talk about everything we know so far about Opus 4.7.

But then I think more importantly, what should you be doing right now to prepare? How do you avoid getting locked into the wrong platform?

So as you expect, Opus 4.7 beats 4.6 at most benchmarks, but it actually does underperform in a few. Unsurprisingly, it absolutely crushes all the coding and knowledge work benchmarks.

On most of these benchmarks, it falls right in between Opus 4.6 and its new mythos model. If you're not familiar, mythos was the model that was deemed too powerful to release because of its cyber security capabilities...

easy for Anthropic to fix without updating the model. However, the update that I think is most relevant, the update that you're most likely to notice and is most likely to impact your decision on how you work with these models, is it Opus 4.7 follows your instructions much more closely, where previous models were much more likely to follow your instructions loosely.

Or guess what you meant, Opus 4.7 follows your instructions literally. If I could speculate a little, I think this is probably the result of how all the model makers are currently training models.

A lot of the rapid advancements we've seen recently are from these model makers realising that if they can verify the output, they can train the models much quicker. This works best when you're training it on coding tasks or anything for which there's a clear benchmark.

But I suspect that training it against these deterministic outcomes is going to make the model less intuitive. Perhaps that's why we're seeing these newer models follow instructions more literally, and why we can expect that trend to continue...

ms work with skills. Well built, robust skills will easily become your business's most important asset, and you can easily move them across any other platforms.