AI budget and costs
Where tokens go, how usage is bounded, and what you pay in each route
Where tokens go
Three kinds of model calls exist, and they are very different in size:
| Call | When | Input | Output (max) | Model |
|---|---|---|---|---|
| Item summary | Once per new item, after fetching | Title + up to 8,000 characters | 800 tokens | Summary model |
| Profile rewrite | Once per digest run, only if there is new feedback | Current profile + new ratings | 800 tokens | Summary model |
| Digest | Once per digest run | Up to 300 item records of ≤ 1,500 characters each, plus profile and feedback | 24,576 tokens | Digest model |
The digest call dominates: it reads every candidate item and writes every section. A busy week with many sources costs more than a quiet one; a source that publishes a lot costs more than one that publishes weekly.
How usage is bounded
There is no spend cap inside the core. What keeps usage bounded are structural limits:
- at most 100 summaries per fetch run (the rest waits for the next run),
- at most 300 items per digest, shared proportionally across sources,
- only items from the last 30 days are digest candidates,
- stored content is capped at 20,000 characters and summary input at 8,000.
Token usage is recorded for the digest call (input and output tokens per digest); summary calls are not metered individually.
Self-hosted: what you pay
The stack itself runs on hardware you already have. The only thing that costs money is the LLM API: at the default models, a few cents a week for the summaries and the digest. With Ollama or another local server it costs nothing. Email is optional and fits Resend's free tier or any SMTP server you already have.
A Claude subscription does not cover API use: the Anthropic provider needs API credits.
Cloud: the AI budget
In the cloud the model calls are included in the subscription. Every account has a monthly AI budget that resets on the first of the month; Settings shows how much of it is used as a percentage under AI budget.
When the budget is used up, new summaries and digests pause until the first. Nothing is lost: sources, topics, preferences and everything already summarised stay as they are, the scheduled digest simply resumes next month. You get one email about it per month. The trial has a smaller budget and a shorter fetch window, so a new account's backlog cannot exhaust it in the first hour.
See Trial and subscription for what the plan includes.