Anthropic just said they're going to limit the ability of the new model to assist with building foundational models.

The ability to assess the user's goal and tune the inference accordingly is there. So, what methods are there to tell if a provider has their thumb on the scale?

Plodding along a longer path to the same goal would burn more tokens without necessarily a decrease in solution quality. Maybe a few more docs.

It'd probably still be worth it to make tokens cheaper. Jevons paradox [1] seems to apply here, where making compute cheaper unlocks use cases that were previously priced out, increasing overall demand.

[1] https://en.wikipedia.org/wiki/Jevons_paradox