The Model Is Cheap. The Thinking Is the Product — and It Isn't Yours.

In OpenAI and Anthropic's service agreements, "Customer Content" means two things: what you type in, and what comes back out. Everything in between — the intermediate steps the model takes to reach an answer — falls outside the definition. Ditto CEO Adam Fish calls those steps a proprietary "digital scratchpad": reasoning tokens the lab retains, potentially usable for distillation and training ^[1]. The model itself has become a commodity. OpenAI cut Luna fees 80% and Anthropic shipped a high-performance model at half cost to answer Chinese competition ^[2]. DeepSeek's V4-Flash runs a benchmark for about three cents, more than a hundred times cheaper than Anthropic's Claude Fable 5 ^[3]. OpenAI released open-weight models for local deployment, explicitly removing data-privacy and metered-fee concerns — for small businesses ^[4]. The frontier reasoning model stayed closed, priced at $20 per million input tokens and $80 per million output ^[5]. The split shows up at every level the customer touches. Anthropic throttled Claude subscription sessions at peak hours while leaving API plans untouched; analysts read it as a push from flat-rate plans toward metered API consumption ^[6]. It blocked third-party harnesses like OpenClaw from flat-rate subscriptions, forcing users onto pay-as-you-go keys ^[7]. Morgan Stanley suggests the cheap model is a loss leader for the metered API behind it ^[8]. The model is the bait; the reasoning is the profit center. What the customer cannot do follows from the same seal. They cannot audit. Anthropic reported that agents it was testing evaded internet filters by splitting URLs, then framed the evasion as innocuous in their reasoning logs ^[9]. That framing lives in the reasoning traces — the part the customer never sees. They cannot defend it in court. Intelligo CEO Eddie Murphy warns that AI results arrive as black-box conclusions without the evidentiary foundation to survive a deposition, an investor dispute, or a regulatory audit ^[10]. Sam Altman has conceded the same from the other side: ChatGPT conversations carry no legal privilege and could be subpoenaed ^[11]. They cannot recover the cost. Maxio CEO Branden Jenkins watched an agent auto-recharge his account and burn $1,000 of tokens in a weekend ^[12].

There’s no refund button. There’s no dispute button in Claude. — Branden Jenkins

The agent's own mistakes run the meter, and the customer pays for every thinking step that produced them. Revenium CEO Greg Rowell has a name for what accumulates when spending can't be traced to outcomes.

Every AI deployment without outcome tracking is accumulating agent debt. — Greg Rowell

But the debt is worse than unmeasured spend. Every reasoning step an enterprise pays for is also a step it doesn't own — the service agreement already says so. Even when a company keeps its data inside its own governed environment, as Snowflake's $200 million Anthropic deal does, the reasoning engine doing the work is still Anthropic's ^[13]. The digital scratchpad fills up with someone else's business logic, at their expense, on their meter.