Death to (and of) Tokenmaxxing
by Ben Letson
Among the Anthropic headlines swirling this week (not the one where the US government pulled a piece of software off of a private company's menu) there was a bullet point that's going to cost AI users a lot of money:
On June 23, we'll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we'll extend the included window.
24 words in Anthropic's Fable 5 release blog reshaped the monetization of LLM usage. To unpack a bit, what Anthropic is doing is allowing the general public to get a sense of what their big hitter model is capable of, then adding in a new layer of how user's buy access. The first taste is free then we settle up as we go.
Anthropic Has Been Edging Toward This
To be fair, Anthropic has been edging towards this for a while. Users get allocated a number of tokens on their Claude subscriptions on rolling 5 hour, 7 day and monthly windows. Any tokens used above those limits get paid for by the token per their API pricing. The vibe is that, post-Openclaw, token consumption is outpacing supply. Plus Anthropic needs to turn on the revenue taps for a clean IPO.
The Fable 5 monetization (assuming it comes back on the market) is an extra layer on top of that. The "best" models (scored using internal "trust me bro" benchmarks) might get locked behind pay-per-token walls now. I think this is the first step in eliminating the subscription model for LLM usage.
The Token Leaderboard Was Always Silly
One of the most bizarre emergent phenomena of the last 18 months is the Token Leaderboard. The idea was to reward (and/or mandate) employee AI usage. Leaving aside that you could prompt for the digits of pi ad infinitum, the fact is that more tokens does not equate to more productivity. In fact, if you ask some of the senior devs getting crushed by 10,000-line code reviews, I think the opposite may be true.
Enter Token Minimax-ing
When the tokenmaxxing lifestyle collides with pay-by-the-token at Fable 5 pricing (2X Opus for reference), I think a lot of companies are going to rethink their AI usage patterns. I suspect we'll end up landing in the world of token minimax-ing (what's the smallest number of tokens I can spend to achieve the best outcome for X task?), which might also lead some folks to rethink how they measure productivity gains driven by AI.
Feels like the end of an era.