Every AI cost dashboard has the same silent assumption: that the price list in its database is still correct. It usually isn’t. Providers change prices constantly — a new model ships, an old one gets cheaper, caching gets its own rate. The moment that happens, a cost tool built on a static price table starts lying, and its historical numbers quietly rot.
Ambral was built on a different premise: pricing is versioned data with an effective date, not a mutable field. This post explains how the pricing intelligence subsystem actually works — because it’s the reason every number we show can be trusted.
Prices are versions, not values
When we record what a model costs, we don’t overwrite the old number. We append a new pricing version: input price, output price, cached-input price, reasoning price, an effective_from date, and a source. When an event comes in, we resolve the version that was in effect at the event’s timestamp — never “the current price.”
That single decision is what makes historical costs correct. A task you ran in March is priced with March’s rates, forever. When Anthropic cuts a price tomorrow, your history doesn’t silently change to match.
Nothing goes live without a human
Automation collects. It doesn’t decide. A collector polls provider pricing pages and APIs, normalizes them into a common schema, and — if it sees a change — files a pending change. That change sits in a review queue until a human approves it into a new version. The system can propose; it can never silently rewrite your economics.
Every version also carries provenance: where the price came from (official page, official API, manual), when it was retrieved, and when it was verified. Trust isn’t a vibe — it’s a field on the record.
Unknown price is never zero
Real stacks use models we haven’t seen yet. The worst thing a cost tool can do is price them at zero, or drop the event. Ambral does neither: an event with an unknown model is accepted and marked unpriced, retained, and backfilled the moment verified pricing exists. Unknown cost is visible, not vanished.
Every number shows its work
Because pricing is versioned and itemized, we can always show the arithmetic: 1.2M input × $2.50 + 42K output × $10.00 = $3.42. A cost without an explanation is just a rumor.
Why this is the hard part
The price data itself is public — anyone can scrape a provider’s pricing page. What’s hard is the pipeline: collecting continuously, normalizing across providers, detecting changes, verifying them, and resolving every event against the right version in time. That’s the thing worth paying for. It’s also why Ambral dogfoods itself: the same engine that prices your agents prices our own cost to serve.