If your tokens run out fast and all your subscriptions are burning up right now, read this carefully.
It sounds dumb, but it is brilliant.
So, I ran a deep research on how much a request and a conversation in LLM chats actually costs. Loosely speaking, Opus 4.8 against Opus 5.5.
I, the idiot, assumed, and I think plenty of people here assume the same, that Opus 4.8, being the old model, costs less than the new one. It is logical, after all: the new one should cost more.
Turns out I am simply an idiot.
In short, the old models cost more.
And would you believe it
I have four Claude subscriptions at the moment, as you remember. And a couple of them still have limits left. Despite the fact that today is Thursday. Which is to say, I just moved to the new model.
So I think: fine, maybe this is only true in Claude. Then I ran a similar research through Codex.
Same thing, in short: 6.1 Sol is the most economical model per request. That is, you remember that Codex and GPT have now cut their limits. But if you work in these models, 6.1 Sol, the freshest ones, you will save the most.
New models are cheaper in tokens than old ones.
Which means every LLM provider is now cutting the number of tokens available in your subscriptions. But they are cutting it, let me remind you, while pushing people to use frontier models.
So somehow they achieve that saving and thereby nudge people towards frontier models. Frontier means the very newest.
In short, just plain move to the new model. And if you do not believe me, run the deep research yourself. You will see that in tokens the frontier model, meaning the very newest model, is the most economical one.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261023.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261023.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)