ADR-0107: FinOps token rule and three amounts

• Status: accepted
• Date: 2026-10-30
• French version: ../ADR-0107-finops-jetons.md

Context

Providers count cached and reasoning tokens differently.

Decision

input_tokens includes cached tokens, output_tokens includes reasoning tokens. Each usage event carries three amounts: estimated, measured, billed (statuses estimated, measured, reconciled).

Consequences

The quota is reserved before the call then committed at the measured amount (429 quota_exceeded).

Evidence

• services/control-plane/src/eai/domains/finops/emit.py
• sql/postgres/V031__finops_usage.sql
• services/control-plane/src/eai/domains/finops/reservation.py
• services/control-plane/tests/test_finops_usage.py