AI Notes: Using Models Keeps Getting Cheaper, Building Them Keeps Getting Pricier

On Monday, October 26, two documents appeared on the same day.

One was the pricing OpenAI gave when it released GPT-6.1 Terra: $1 per million input tokens and $5 per million output tokens, half the price of the previous GPT-5.6 Terra and of its own stronger GPT-6.1 Sol. OpenAI also said that on one benchmark it matches the flagship GPT-6 Astra at a tenth of the price per token.

The other was the S-1 Anthropic filed publicly with the SEC, planning to list on the Nasdaq under the ticker ANT. The filing lists long-term compute commitments to three cloud providers, Google, Amazon and Microsoft, totaling about $269 billion, and states that this money must be paid "regardless of usage."

One says using models keeps getting cheaper; the other says building them keeps getting more expensive. The two don't contradict each other. They come from the same structure.

Why prices keep falling

On the surface, falling token prices are about competition. Underneath, there are three forces.

The first is distillation and better architectures. Once a frontier model exists, the same capability can be squeezed into smaller, faster models; a mid-tier model like Terra is essentially the last flagship's capability put into mass production. The second is engineering on the inference side. Caching, batching and specialized chips keep pushing down the marginal cost of each token. Terra's cached input price is $0.05 per million, just 5% of regular input, and that discount alone tells you about the cost structure: repeated context costs almost nothing. The third is that open weights put a ceiling on prices.

On October 29, Mistral put the weights of Large 4 on Hugging Face: about 1.05 trillion total parameters, with about 52 billion active per token. Its hosted API is priced at $0.70 per million input tokens and $2.10 per million output. More importantly, any organization with a single eight-GPU B200-class server can deploy it itself. As long as the market has an option that is close in capability and can be run in-house, closed models will struggle to hold high prices in the middle tier.

So for models at mid-tier and below, prices are getting ever closer to inference cost plus a bit of margin. This layer is deflating.

The other set of books in the S-1

Anthropic's S-1 is the first time a frontier lab has laid its full accounts open to the public.

Revenue is growing fast: $16.3 billion in the first half of 2026, against just $1.2 billion a year earlier. But what really shapes the company is the cost side. The filing says cost of revenue is "almost entirely" related to compute and infrastructure; fixed or minimum-guaranteed compute commitments come to about $265.6 billion, with terms of 1 to 20 years; counting equipment leases and forward capacity not yet priced, total commitments are between $542 billion and $567 billion.

There is a maturity mismatch here. Revenue is collected one token at a time, and customers can switch models or cut usage whenever they like; costs are signed by the year, and once signed they must be paid. Put the flexibility of revenue and the rigidity of costs in the same statement, and margins will swing up and down with the pace at which new compute comes online. The S-1 says as much itself.

More worth watching is what the commitments are for. Outsiders can see compute falling as a share of revenue, but not how much of these commitments goes to training the next generation of models and how much to serving existing customers. Compute that serves customers gets diluted by revenue; compute for training is prepayment for the next round of competition, and it produces no revenue in the current period.

Why a training pause made the risk factors

The S-1's risk factors include a disclosure: Anthropic once paused pre-training of its next-generation frontier foundation model for more than a month to strengthen protection against theft of model weights and intellectual property. The filing also warns that similar pauses could delay revenue and leave the company behind its rivals.

Seen through the cost structure, this has a very concrete meaning. Compute is signed by the year, so during a pause of more than a month the compute doesn't stop billing. It can be shifted to serving customers, but the timeline for the next model slips. In an industry that holds up pricing with each new generation of models, the time between generations is the timing of cash flow.

Gemini 4 Carbon, which Google released the same week, is an example from another angle. It is Google's most capable model so far, yet it is first open only to a subset of cyber defenders, and will be opened to developers and enterprises after the guardrails have been iterated. The published starting price is $2 per million input tokens and $10 per million output. At the very frontier, the release schedule is starting to be constrained by safety, not just driven by competition. Safety spending has become part of the cost of building, and it directly affects the day revenue starts to be recognized.

One side's commitment is another side's capex

The $269 billion Anthropic owes three cloud providers is, on the cloud providers' side, backlog, and the basis for expanding their data centers. This quarter, free cash flow at both Alphabet and Amazon turned negative, as it did in the quarterly report Meta released a little earlier.

Connect the chain: model companies commit to long-term purchases from cloud providers; cloud providers raise capex on that basis; the capex first eats into their own operating cash flow and is then slowly recouped through the model companies' payments; and those payments ultimately come from token revenue at $1 or $5 per million. The end of the chain is cutting prices while the front end is doubling down, and every link in between is betting on the same thing: that usage grows faster than unit prices fall.

Whether usage will keep up is, so far, a matter of mixed evidence. The same week, DoorDash CEO Tony Xu said on the earnings call that orders from third-party AI agents came to "a few hundred in total," while DoorDash handled 993 million orders in the third quarter alone. Agents are considered the most token-hungry use, and Anthropic's S-1 also notes that agentic products are structurally more compute-intensive. Yet in concrete consumer settings, the real transactions agents bring in are still very few.

Which layer value lands in

Put the two ends together and you can roughly see where profits will flow.

The mid-tier is deflating and open weights set a ceiling, so profits in this layer will get thinner and thinner, left to whoever can push unit costs lowest. The very frontier can still set prices, but it has to carry ever higher training costs and ever longer safety cycles, and lock in compute ahead of time with long-term commitments. Cloud providers, holding the compute and the entry point to customer budgets, receive relatively certain cash over the commitment period, at the price of draining their own cash flow first.

That is what the IPO disclosure means. These accounts used to be hidden in private financing terms and partnership agreements. Now one of these companies has written its cost structure, commitment terms and risks into a document it answers to public investors for. From here on, when the market prices these companies, it will lean more on these books, not just the slope of the revenue curve.