Your reply is not what you are paying for.

The message this page is costing

Rates below are read from the current catalog. These token counts are not — they are an assumption about a long session, stated so you can substitute your own.

Input

6K tokens

Card, persona, lore book and the history still in the window.

Output

300 tokens

Two or three paragraphs of reply.

Ratio

20:1

Input to output. This is the number that drives everything below.

Where the money goes in one message

Priced on Llama 3.3 70B Instruct, at its current catalog rates.

Input is 95% of the message. The practical consequence is that shopping on output price — the instinct most people bring from code-generation workloads — optimises the smaller half of your bill. Compare input rates first.

And regeneration is cheaper than it feels. Pressing regenerate re-sends the same prefix and generates a new reply, so it costs about what the original message cost — not double, and nothing like restarting the thread. The expensive action in roleplay is growing the prefix, not producing text.

A 1,000-message session

The same assumed message, costed across the models on the roleplay facet, at current catalog rates.

Every number above is arithmetic on published rates, so you can check it: input tokens times the input rate, plus output tokens times the output rate. Nothing is estimated and nothing is averaged over other people’s usage. Redo it with your own token counts from the catalog or /pricing.

What actually reduces the number

In order of how much they move it, for this workload.

1. Trim the prefix, not the reply. A lore book entry that never comes up still costs you on every single turn. Cutting a thousand tokens of dead weight from the card is worth more than any setting on the output side.
2. Compare input rates first. Two models with the same headline price can differ severalfold on the side you are actually buying.
3. Watch the window, not just the price. A thread that fits at message 20 may not at message 200, and the failure arrives as a 400 rather than a gradual slowdown. Context windows are on every model page and on /models/long-context.

Prompt caching is not in these numbers. A mostly-static prefix is exactly the shape that cache reads are meant to make cheaper, and it is the obvious next lever for this workload. We are not claiming a cached rate here because the numbers on this page are the ones the catalog currently publishes, and putting a hypothetical discount next to real rates would make the whole table untrustworthy.