What this costs to run
Every time a worker runs, one line gets written down: which model it used and how much it consumed. Nothing else is stored — no prompts, no outputs, no business figures. The money is worked out when you look at it, from a single price list, so a price change re-prices the whole history at once.
What has been recorded
A hook writes one line every time a worker finishes. Until it is installed on this machine, the bill is unknown rather than zero — see what needs you on the front page.
vault-schema/AI-HQ/Runs/routing.mdThe meter never blocks a run. If it fails it fails quietly, because a broken cost meter that stops the org from working is worse than no cost meter at all.
The price listwhat a million tokens costs, as of 2026-06-24
| Model | Fresh input | Reused input | Output |
|---|---|---|---|
| haiku-4.5 | $1.00 | $0.10 | $5.00 |
| sonnet-5 | $2.00 | $0.20 | $10.00 |
| opus-5 | $5.00 | $0.50 | $25.00 |
| opus-5-fast | $10.00 | $1.00 | $50.00 |
| fable-5 | $10.00 | $1.00 | $50.00 |
Look at the neon column: reused input on the most expensive model costs less than fresh input on the cheapest one. That is why reuse is fixed before anyone argues about which model to use. Anthropic first-party API pricing, via the bundled claude-api skill's cached model table. Re-verify at https://anthropic.com/pricing before any figure reaches a human.