AtoAxisWhat is recorded
A usage event contains identifiers, provider and model names, an event time and numeric token counts. Input, output, cache-read and cache-write tokens remain separate because providers can price each class differently.
{
"provider": "example-provider",
"model": "example-model",
"input_tokens": 1000000,
"output_tokens": 500000,
"token_source": "reported"
}How the estimate is calculated
For every token class that the provider reports, AtoAxis multiplies the count by that class rate per million tokens, divides by 1,000,000 and adds the result in micro-USD. A class that is absent does not block the estimate; a reported class without a known rate makes the event cost unknown so a partial total cannot look complete.
Example rates (illustrative only):
input 1,000,000 × $3.00 / 1M = $3.00
output 500,000 × $15.00 / 1M = $7.50
estimated total = $10.50 = 10,500,000 micro-USDProvider, model and effective date
The lookup matches both provider and model. It then selects the price window whose effective-from and effective-to dates contain the event timestamp. A later price change therefore does not silently rewrite an older event, and the pricing version can be inspected with the result.
- The same model label at two providers is treated as two different price records.
- A provider or model with no matching price remains unpriced.
- A cache class with no rate makes cost unknown when that class is present.
- Historical events keep the rate effective at their own event time.
Why unknown is not zero
A zero could mean free usage, missing usage, an unpriced provider or a genuinely zero count. Those cases are not equivalent. AtoAxis reports token usage when it has it and leaves estimated cost unavailable when the price evidence is incomplete, with coverage explaining how much of the range is priced.
Project totals and limitations
- Usage is grouped by project when the session provides an unambiguous project context.
- A session that spans several projects stays Unattributed instead of being copied into every project.
- An estimate is not an invoice: negotiated rates, subscriptions and provider-side adjustments may be outside the visible data.
- Prompt and response text are not required and are not stored in the usage event.
Review and verification
Use the AI cost tracking page for supported agents, then compare the dashboard token total, priced-token coverage and pricing version for the same range. Review owner: AtoAxis product team. Last reviewed: 20 September 2026.