Anecdotally my work is already pivoting from "use as many tokens as possible" to "try and use a dumber model if you can because the best ones are too expensive".