The spec is at v1.0.0 under an Apache 2.0 license. Each record holds the raw counts that drive cost, such as tokens, tool calls, seconds of compute, plus the customer, feature and environment the cost belongs to. Spawned agents reuse the parent's run_id and link back through parent_span_id, so a full execution tree adds up to one run. Chargebee has published Python and TypeScript sinks that send records to its usage-based billing. Mastra, Merge and others have adopted it. AUDR records usage. It does not bill. Teams on Stripe need a sink that drains AUDR events into Stripe's meters, and records can also go to a file or an existing OpenTelemetry pipeline. The pitch is about margins. A Chargebee senior director said AI companies can chart revenue but not margin, and that "the cost of a single agent run can be 10X comparatively." For founders pricing agents by outcome or usage, a common record that links each run to a customer makes cost-to-serve something they can measure instead of estimate.