The Labs Are Giving Away the Model and Sealing the Thinking

Start with the fine print. In the service agreements that govern most enterprise use of OpenAI's and Anthropic's models, "Customer Content" is defined as two things: the input a customer sends and the output the model returns. Everything between them is not customer content at all. The thousands of intermediate reasoning steps a model runs to turn input into output are, by the plain reading of the contract, the lab's. Ditto CEO Adam Fish, whose company stores AI output, argues the labs lean on exactly this gap: customers pay for reasoning that runs on the lab's machines, gets billed to the customer, and belongs to no one the customer can reach ^[1]. That clause is the whole mechanism. The reason a buyer cannot audit the thinking it pays for is not a technical limit. It is a definition. Every major lab now runs its own variation. OpenAI is the most direct: it withholds the raw chain of thought to protect a competitive advantage ^[1]. And when 17 news organizations demanded records in a copyright suit, OpenAI spent two years claiming it could not search its own training data, then acknowledged it had deleted billions of conversation logs. The publishers are now seeking sanctions ^[2]. Anthropic hides the same layer by billing it. Reasoning tokens arrive encrypted and invisible, so a customer is charged for processing it can never inspect ^[1]. When the open-source tool OpenClaw tried to reach Claude under flat-rate subscriptions, Anthropic blocked it, leaving users with pay-per-token access or the lab's own tools. The tool's creator read the move plainly.

Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source. — Peter Steinberger

The same lab now deepens its hold on the other side of the exchange. Anthropic's enterprise Claude Desktop plugs into Microsoft 365 and DOD endpoints, which puts the lab in view of both a company's data and the model's reasoning while the company sees only the final sentence ^[3]. Google runs the third variation: gating. Its most powerful reasoning models are closed and tiered. Gemini 2.5 Deep Think gave the broad public a speed-optimized version while the full reasoning variant went to a hand-picked set of mathematicians and academics ^[4]. Gemini 3 Deep Think, launched in February, produces gold-medal-level answers but hands users only the final one, and access is limited to Google AI Ultra subscribers and select enterprises ^[5]. Now look at what these same labs are giving away. Prices were cut by as much as 80% to answer Chinese open-source competition, and OpenAI released small open-weight models anyone can run locally. Anthropic and OpenAI signed sovereign agreements with India and the UK that let customer data sit on domestic servers. Morgan Stanley's analysts read cheap and open model access as a loss leader: the giveaway that pulls in spending on storage, security, and managed APIs, the profitable layer ^[6]. That is an analyst's observation, not a lab's stated plan, but it matches the shape of everything above. The model becomes the commodity. The reasoning around it becomes the thing you pay for without ever seeing. The frameworks meant to hold this to account are all aimed one layer too high. Europe's transparency rules cover inputs and outputs: labeling AI-generated content, telling people when they are talking to a machine, disclosing training data. Nothing requires a lab to expose the steps between input and output ^[7]. The sovereign-AI movement is about where data sits, not how it thinks. The UK and India agreements specify residency and audit trails for input and output, and are silent on intermediate reasoning ^[8][9]. Even the people inside the labs see the gap. Forty researchers from OpenAI, DeepMind, Meta, and Anthropic, including Ilya Sutskever and Geoffrey Hinton, warned last year that reasoning models' inner workings are going dark, and cautioned that even the transparency they ask for is not guaranteed to survive as the models advance ^[10]. Nvidia put the final point on it in August. Its researchers showed that the harness surrounding a model matters more for long-horizon reasoning than the model itself, lifting Claude Opus 5 from a 30% score to a perfect one on the ARC-AGI-3 benchmark ^[11]. That finding inverts the whole transaction. If the value in reasoning lives in the surrounding harness, not the weights, then the labs are giving away the part that matters less and enclosing the part that matters more. Regulators are checking the labels. Sovereign governments are guarding where the data sleeps. Open-weight challengers are freeing the weights. None of it reaches the middle, where the thinking happens and where no one but the lab can look.