Garry Tan's AI Distillation Push: Pricing Impact

Y Combinator CEO Garry Tan wants U.S. open-weight AI labs to be allowed to distill capabilities from American frontier models through legitimate access. In a CNBC interview published September 11, Tan answered “I would do nothing” when asked how the U.S. should respond to Chinese labs distilling American models, then suggested an “American distillation regime.”

Tan clarified to TechCrunch that he is not advocating stolen credentials, fake accounts, or covert access. His proposal is a legal, “front door” route that would let smaller U.S. labs use frontier-model outputs to train open-weight alternatives.

The immediate pricing verdict: nothing on a public rate card changed. The live table, chart, and calculator above compare current closed and hosted-open routes; they do not estimate the cost of a model that has not been announced.

What Tan proposed—and what he did not

Question	Tan’s position	What remains unresolved
Should regulators broadly stop distillation?	No; Tan told CNBC he would “do nothing”	Whether Congress or agencies will define an authorized regime
Should U.S. open-weight labs distill frontier models?	Yes, through legitimate customer access	Which outputs, volumes, models, and uses would be permitted
Are stolen credentials acceptable?	No; TechCrunch says Tan explicitly excluded them	How providers would distinguish research from extraction at scale
Did API terms or prices change?	No change was announced	Providers could revise contracts, access controls, or licensing later

Distillation uses a more capable “teacher” model to produce examples for training a smaller “student” model. It is a standard machine-learning technique when authorized. The dispute is over permission, scale, provenance, and whether API terms can prohibit using outputs to build competing models.

Why Anthropic disagrees

Tan’s comments follow Anthropic’s September threat report, which says the company identified and disrupted additional distillation attacks involving seven labs based in China. Anthropic distinguishes legitimate distillation from covert, industrial-scale extraction that uses fake accounts, stolen payment methods, compromised credentials, or resellers.

Anthropic says the campaigns targeted valuable capabilities including agentic tool use, coding, data analysis, and reasoning. It argues that unauthorized labs can imitate those capabilities with less time and compute than independent development. CEO Dario Amodei has separately called for stopping industrial-scale distillation while opposing a blanket ban on open-weight models.

Tan’s counterargument is economic and political: frontier labs trained on broad public data, customers should have more freedom over outputs they paid to generate, and one dominant proprietary provider would be a dangerous outcome. That leaves a real policy gap between ordinary model training and fraudulent extraction.

Pricing impact: competition later, no discount today

Authorized distillation could reduce the cost of building capable U.S. open-weight models. If those models are released and perform well, third-party hosts could compete on inference rates while enterprises gain self-hosting and private-deployment options.

None of that produces a measurable saving today. There is no named student model, training dataset, license, hardware profile, release date, or hosted price attached to Tan’s proposal. Buyers should compare current routes on cost per accepted task instead of assigning a speculative discount to “open” or “distilled.”

Open weights also do not mean free production inference. GPU capacity, utilization, redundancy, storage, networking, monitoring, security, and engineering remain part of the bill. See our Meta Llama API pricing guide for the hosting economics, then compare Anthropic pricing, OpenAI pricing, and Together AI pricing in the token calculator.

Labs decision: policy is not a model route

Tan’s proposal does not enter the Cost-per-Task Leaderboard. It provides no exact student model, downloadable artifact, callable endpoint, fixed task set, training recipe, acceptance grader, token ledger, or public deployment price.

We recorded a policy and availability blocker. A future U.S. open-weight release becomes eligible when the exact artifact or route can be reproduced with a defensible current cost basis and complete token, latency, retry, correctness, and billed-cost evidence.

Who benefits—and who loses

U.S. open-weight startups would benefit most if an authorized regime gave them a predictable way to use frontier outputs for training. Developers and enterprises could gain more capable models with downloadable weights, multiple hosting routes, and less dependence on one API vendor.

Frontier labs could lose some capability advantage and pricing power. They would also bear the cost of serving training-scale traffic unless access terms and fees explicitly account for it. A clear paid licensing route could turn that burden into revenue, but Tan did not outline one.

Chinese open-weight labs could lose a relative advantage if U.S. competitors gain comparable training access. Customers lose if the argument instead leads providers to tighten onboarding, rate limits, monitoring, or output-use terms for legitimate high-volume workloads.

What developers should do now

1. Keep current API budgets unchanged; Tan announced a policy preference, not a product or price.
2. Review provider terms before using outputs for training, fine-tuning, synthetic data, or model evaluation.
3. Separate authorized research from production accounts and document account ownership, purpose, and traffic patterns.
4. Benchmark closed, hosted-open, and self-hosted routes on accepted-task cost, including retries and operations.
5. Keep a fallback provider in case distillation enforcement increases account reviews or access friction.

For background on the dispute, read our earlier Anthropic–Alibaba distillation pricing analysis.

Teams that want to test a managed open-model route can compare Novita’s hosted catalog. Verify the exact model, license, region, context limit, and live rate before moving production traffic.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Bottom line

Tan is proposing a U.S. answer to open-weight competition: let smaller American labs legally learn from frontier APIs instead of trying to ban distillation broadly. Anthropic’s evidence shows why providers draw a hard line around covert extraction, but the rules for authorized training remain unsettled.

For buyers, this is a competition signal—not a price cut. Use today’s published rates until an actual U.S. distilled model, license, and hosting price exist.

Sources: TechCrunch’s report and clarification from Tan, CNBC’s interview report, Anthropic’s September threat intelligence report, Anthropic’s position on open-weight models, and AI Pricing Guru’s live pricing dataset. Sources checked September 13, 2026.