Use IA4AI
How costs are counted
Where each cost comes from, what "measured" and "estimated" mean, and why an unknown price is never shown as free.
#Where a cost comes from
| Source | How the cost is found | Status |
|---|---|---|
| The gateway | The provider's own reply gives the tokens (uncached input, cached input, cache writes, output); IA4AI prices them from its model catalog at the time of the call | measured tokens, catalog price |
| OpenTelemetry and direct ingest | The token counts your agent reports, priced the same way | as reported |
| Claude Code telemetry | Claude Code's own estimate of the API-priced value of the work | Claude Code's estimate, never a bill |
| A provider bill you upload | The provider's own figures, day by day | measured |
Input is always counted as the whole prompt, the cached part included, so a provider that reports only the uncached part is not undercounted.
#The model catalog
Prices come from public price feeds, checked weekly against a second source. A change to a price is proposed and
approved by a person before it is used; nothing in the catalog changes by itself. You can read the catalog, with each
model's prices, capabilities and how it did in your own verification runs, at
GET /catalog.
#Unknown is not free
A call for a model the catalog does not know yet is unpriced, never $0:
- Reports carry
unpriced_calls,unpriced_tokensandunpriced_modelsbeside the totals, and the totals say they exclude those calls. A screen reads, for example, 12 calls on a new model have no price yet: this total is understated. - As soon as the catalog learns the price, unpriced calls are priced. This happens after every catalog update and in an hourly sweep.
- A call that already had a price keeps it, even if the price changes later: costs stay at the price of their time.
- Repricing is on the record: the changed records are stored again with a new evidence entry, and one entry says how many calls changed and for which models.
A model the catalog lists as free (a local model, say) is priced at zero on purpose, and shown as such.
#Check the totals against your bill
Upload your provider's usage export (CSV) on the Spend screen, or with
POST /workspace/provider-bill. The bill check compares IA4AI's totals
with the bill day by day. Only the totals per day and model are kept, never the file. A small gap is normal: calls your agents made outside IA4AI, calls cut off before their
usage arrived, and the edges of the period all show up there.
#Savings
A saving is stated only when it has been verified on your own traffic: a cheaper model replayed on your real inputs and shown to answer the same way. Until then a cheaper model is a candidate to test, with no saving attached. See Capture, golden sets and verification.
#Your own hardware
The Own hardware screen (or GET /report/alternative-plan, and the alternative_plan
MCP tool) sets your bill against running open models on your own servers, in three tiers:
| Tier | Takes | Model | Hardware options |
|---|---|---|---|
| Good | Easy work | Qwen3-Coder-Next, or Qwen3.5-122B on a desk box | A desk box (Mac Studio M5 Max, DGX Spark, Ryzen AI Max+), or a server with one 96 GB card, bought or rented |
| Pro | Moderate work | GLM-5.2 | A server with 8 H200 GPUs, bought or rented |
| Premium | Hard work, and work with no difficulty label | Kimi K3 | A server with 8 B300 GPUs, or two with 16 B200 GPUs, bought or rented |
Each call goes to the lowest tier that could take it, from its difficulty and its length. For your volume the screen sizes every hardware option, takes the cheapest to buy and the cheapest to rent, and shows what each mix costs a month beside what the same calls cost now. Hardware prices and speeds are published figures, each with its source and date, checked every week.
Two choices are yours, on the same screen:
- Privacy. No preference suggests whatever costs least. Prefer our own hardware suggests running it yourselves when that costs about the same (within 20%). Data must stay in-house rules the API out. The same choice limits a verification replay: with either of the last two, your content is never sent anywhere it does not already go, and with Data must stay in-house only models on your own hardware are tried.
- Seat prices. What you pay a month per seat of each chat app. Your chat seats then count in the bill, and licensed seats nobody used show what they cost.
As everywhere else, no mix is a saving until your own calls have been replayed on the open model and it answered the same way.

