90% Cheaper Tokens, One Legal Moat: OpenAI's Anti-Distillation Terms Are a Pricing Power Play

The new restriction is drafted to close every route by which a rival could turn paid OpenAI output into a competitor. The stated reason is safety: training another model on OpenAI outputs can transfer capabilities without carrying the safeguards over. The commercial effect is that a paid inference bill stops being a legal raw-material source. Critically, ordinary enterprise use is untouched — research, software development, model evaluation, benchmarking and product development stay permitted within boundaries, and existing permitted exceptions under the services agreement still apply. The clause binds model builders, not buyers.

The October policy is the fourth public step in seven months, and each step raised the cost of the same behaviour. Anthropic first named the labs and the mechanism in February 2026, then named a larger cloud customer in June, then the US government named six labs in September, and OpenAI published its own campaign account days later. By October, Anthropic could report that its technical countermeasures were working — data harvested in distillation attacks down about 90% to 99%, and reasoning tokens extracted by distillers 10 to 100 times lower. Once the technical leak is plugged, the legal terms become the binding constraint.

The price action is not generosity. OpenAI's entry tier fell 90% in eight weeks and its frontier tier fell 60%, while Anthropic moved its own tiers to the same price points. Two labs competing on price alone end up with no moat at all, which is exactly why the surviving competitor has to be something contractual. OpenAI's own framing supports the link: revenue and compute scaled almost in lockstep from 2023 to 2025, with ARR moving from $2 billion to $6 billion to more than $20 billion while compute went from 0.2GW to 0.6GW to 1.9GW. The company's run rate then passed $40 billion in mid-August and sat near $50 billion at the end of September. Cheap tokens buy share; the legal fence keeps the share.

That is the pricing power argument. Anthropic now captures about 40% of enterprise LLM spend, with OpenAI at 27% and Alphabet at 21%, together roughly 88% of the market, inside an enterprise generative AI budget that reached $37 billion in 2025. OpenAI's share has fallen from about 50% in 2023. A vendor losing enterprise share cannot win on price, because the buyer who wants the cheapest token is also the buyer most likely to route around a vendor's terms. OpenAI is defending price with paper, not with capability.

The resale approval requirement is the piece that reaches beyond the labs. Anthropic's own investigation found that the dominant access route was not stolen keys but commercial proxy services reselling model access at scale, including one network running more than 20,000 fraudulent accounts simultaneously. OpenAI's answer is to require that any operator letting third parties reach OpenAI models obtain prior review and approval. That lands directly on a layer that just repriced: Stripe agreed to acquire OpenRouter for more than $7 billion, with reported terms of about $7.5 billion including $1.5 billion for the founders. A gateway business cannot be valued on traffic alone once access becomes permissioned.

Here is the part the price-war narrative misses. If a rival cannot distil frontier reasoning, it must train it, and training is the most compute-hungry activity in the industry. Enforcement therefore shifts demand rather than removing it. The evidence is already in the reported numbers: NVIDIA grew revenue 106% year over year to $96.2 billion in the quarter ended July 26, 2026, with data center revenue of $89.0 billion, up 117%, and gross margin widening to 75.0% from 72.4% a year earlier. Token prices fell 90% while the silicon bill rose 117%. That is the signature of a demand shift rather than a deflation.

Two second-order readings follow. First, the capex is being funded from the same revenue pool that the price cuts are shrinking, so free cash flow compression is the visible cost of defending share: Amazon spent $131.8 billion in FY2025 and kept $7.7 billion of free cash flow. Second, Broadcom shows where the overflow goes — revenue of $29.6 billion in the quarter ended August 2, 2026 against $16.0 billion a year earlier, on a 54% operating margin, as hyperscalers built in-house silicon rather than pay for merchant GPUs. The compute bill for the training that replaces distillation is being paid at the margin, not by the frontier labs alone.

Downstream, the effects split across three very different balance sheets. In China, the named labs are well capitalised rather than broken: Moonshot AI raised $3.5 billion in July 2026 at a reported $35 billion valuation, and its Kimi K3 licence already seeks revenue share of up to 30% from large commercial users. That is a company building a business on the back of distilled capability, and the new terms put that capability's provenance in writing. Alibaba's June quarter shows what that costs: revenue up 8.3% to CNY 268.2 billion, but operating income down 44.9% to CNY 19.3 billion from CNY 35.0 billion a year earlier, as AI infrastructure and proprietary chip investment absorbed the gain.

In the days to quarters, the clause is a trial run rather than a settled regime. The first evidence will be procedural: whether OpenAI actually denies or conditions a gateway approval, and whether enterprise buyers push back on the reseller restriction at procurement. Two dated catalysts sit close: the promotional GPT-5.6 Sol rate expires on November 21, 2026, one day after the new terms take effect, and the guidance built into that promo will show whether OpenAI needs the fence to hold price. Watch the reseller refusals, not the press releases.

The evidence separates cleanly into three tiers. Facts: OpenAI's policy text and effective date, the enforcement disclosures from Anthropic, OpenAI and CISA, the published price schedules, and the reported financials of every company named here. Inferences: that the reseller clause is aimed at protecting price rather than model quality, and that the training bans push spending upstream. Speculation: that US policy will ratify the barrier into something durable. The investable conclusion does not depend on the speculation — it depends on enforcement actually binding, and on rivals training rather than distilling.