Per token is coming

by Asim Hussain · 26 August 2026

▶ reveal tech

I built a cost report into my AI harness. It runs after every spec and tells me what the feature cost. One came out at $180 at list rates. I didn’t pay that, I’m on a subscription. But the all-you-can-eat buffet on AI is going to end, and when it does, these are the rates. So act as if you’re paying per token now. The bonus is that cost is the best proxy for environmental impact, so spending fewer tokens is also the greener thing to do.

AI-distilled from the episode

Jump into the video

Not every token is the same

Input is what you put in. Output is what it generates, and it’s also how it thinks, and it costs five times more because it takes more compute. More compute, more cost, more environmental impact.

Opus 5.0 list rates, per million tokens ↻ click to replay

The real money is the whole conversation

The cost compounds, and compounding is something I’ve always struggled with. It’s probably why I never have any money. First turn: your prompt goes in, small, and the answer comes out, large. You respond, and everything from the first turn, what you put in and what it gave back, gets sent through again as the input for the second. Then again for the third. Your first message might cost five cents. By the tenth turn each extra message is costing twenty.

ten turns of one chat, what each one costs ↻ click to replay

That’s why the report added up so fast. Not one big message. Hundreds, each dragging the whole chat behind it. So kill your long chats. Start a fresh one with no baggage, or run /compact, which throws the old baggage away and everything after it rebuilds from a smaller base.

/compact at turn five ↻ click to replay

Caching is what saves you

A long chat should be a disaster, and it would be if it weren’t for caching. Your input and output get cached for the next turn, and a cache read is ten times cheaper than fresh input. In a 200,000-token conversation maybe the last 190,000 were cached. In that $180 report, 97% of the tokens were cache reads. But go idle for an hour on Claude Code and the cache evicts, so the next turn reloads the lot at full price. Stay active.

the same chat, with the cache underneath every turn ↻ click to replay

Three rules

Say “be brief in your responses”. Keep your chats small. And if you can’t keep them small, at least stay active.

Treat every token like you’re paying for it. Because soon enough, you will be.