A quick update before today's article on the AI space.

Anthropic has released Opus 5. It performs on par with Fable 5 on several benchmark scores while being priced the same as Opus 4.8, so naturally the AI community has been loving this release.

I've been testing it since launch and it has genuinely impressed me. The reasoning especially. It's precise and it stays consistent.

But for the past year, the advice across the AI world has been the same: if a model gets something wrong, let it think longer. Turn reasoning on, increase the thinking budget, and wait a little longer.

Now, five researchers from the University of Trento, Fondazione Bruno Kessler, and Toyota Motor Europe decided to test where that advice starts to break down.

They asked a fascinating question:

Once a model has already reached the correct answer, does giving it more time to reason make the answer better or can it actually lead the model away from the correct solution?

To find out, they split each reasoning trace into individual steps and forced the model to produce an answer from every partial reasoning path. This allowed them to identify the earliest point at which the model had already arrived at the correct answer.