AtlasBurn
DocsPlatformIntegrationsOperationsReference

Accuracy

Summary

AtlasBurn's cost numbers are derived from provider-reported token counts and the pricing registry. We do not publish a fixed accuracy percentage today — the streaming-usage capture path changed materially in I3 (June 2026) and any pre-I3 figure would overstate accuracy. This page documents the methodology so you can reconcile against your own invoices.

Methodology over headline numbers

Earlier docs cited 94.2% against provider invoices. That number was measured before OpenAI streaming responses reported usage via stream_options.include_usage — pre-I3 streamed calls logged $0 tokens, so the original sample was biased. We removed the figure and are re-deriving it against current invoices.

How a cost is produced

  1. The proxy or SDK observes the upstream response.
  2. Token counts are taken from the provider's usage object. For OpenAI streams, the SDK and proxy inject stream_options.include_usage: true so the final SSE frame carries real counts.
  3. If no usage is emitted (legacy stream, malformed response, or non-supported provider), tokens are estimated from request size and the event is flagged est/~est in the ledger.
  4. The pricing registry maps (provider, model) to per-1M-token input/output rates. Unrecognized models fall back to FALLBACK_ECONOMICS (conservative rates).

Variance sources

SourceImpactMitigation
Estimated tokens (no usage frame)Most material — event marked estUse a client that supports include_usage, or route via the proxy
Context-cache / batch discountsDiscount visible only on the invoiceReconcile monthly against the invoice
Provider-side roundingSub-cent variance per callAccepted
Fallback pricingConservative overestimate on unknown modelsAdd the model to PROVIDER_REGISTRY

Reconciliation

Reconciliation aligns the Forensic Ledger with the provider invoice line items. Today this is manual via the dashboard; an automated Reconciliation API is planned.

What this means in practice

Treat AtlasBurn numbers as operationally accurate — good enough for guardrails, burn forecasting, and feature attribution. For financial reporting, reconcile against the upstream invoice.

Next Steps