The useful unit of AI work is not a token, a seat, or a burst of activity. It is a task that made it all the way through the quality bar. That is the core idea in OpenAI's July 17 scorecard for the AI age: count useful work, count the full cost of a successful task, measure dependability, then ask whether the economics improve as work scales.
That is a better question than "which meter looks cheapest?" A low token price can still produce an expensive outcome if the task needs retries, human edits, or a rescue pass. The scorecard explicitly puts review and rework back into the equation. For Hugin, that is a real primary record of how OpenAI says it wants customers to measure value. It is not an audit of anyone's actual workflow, and it is not a service-level agreement.
A score is not a control plane
The second record matters because successful work is only useful if the agent was allowed to do the work in a legible way. In its May 8 Codex deployment write-up, OpenAI describes a separate control plane: sandbox boundaries, approval policy, network rules, credential handling, command rules, and logs that can explain what the agent tried to do.
Those are not interchangeable claims. A scorecard asks whether a completed task was valuable. A control plane asks what data and systems an agent could touch, where it had to stop, and whether somebody can reconstruct the decision later. OpenAI's document is strong evidence of the controls it says it uses in its own deployment. It is not evidence that every user, plan, workspace, or local configuration inherits every one of those controls.
A reporting promise is a third kind of receipt
Anthropic's July 9 "Inviting hard questions" announcement belongs on neither of those first two lanes. It asks the public for difficult questions about AI and says Anthropic will publicly track and report the actions it takes in response. That is a dated commitment. It becomes evidence of results only when later reports actually arrive and can be compared with the promise.
This is the small discipline the AI desk is trying to make normal: a framework is not an audit; a deployment description is not a universal product guarantee; a promise to report is not the report. All three can be useful. They just do different jobs.
What changed in the case file
- AI Release Receipts now carries OpenAI's May 8 control-plane record and July 17 useful-work scorecard as separate provider-authored sources.
- The case file keeps Anthropic's July 9 public-reporting commitment in its own future-facing lane.
- Today's companion journal logs one observed capacity reset as account evidence only — a ledger entry, never a promised cadence.