Meta's No-Training Promise Lasted Fifteen Months

Meta replaced its Llama API Terms of Service with a new Meta Model API Terms of Service, released July 8, 2026. The old terms made a blanket promise: Meta would not use your Content to train any AI models. The new terms keep that promise for paid traffic and reverse it for unpaid traffic. So this is not simply a new pricing tier. A protection that used to cover everyone has been carved in two and sold back to the paid side. On the free service, your inputs and outputs can be used to train and improve Meta’s AI. On the paid service, Meta says they will not. There is no switch inside the free tier to hold specific traffic back. If you need that protection, Meta’s answer is to pay.

In human terms: A three person studio wires Meta’s model into a client tool to see how it performs. They run the trial on the free tier, because that is what free tiers are for. Every prompt they send, including the client’s brief and the sample assets pasted in to test it, is now eligible to train Meta’s models. Nothing warned them at the keyboard. The only way to have kept that material out was to have been on the paid service before they started typing.

Why this matters: For years the practical difference between a free and a paid API was cost, speed, and rate limits. Meta has added a new axis: confidentiality. Whether your inputs and outputs feed someone else’s model is now a billing decision you make before the first request, not a setting you toggle afterward. For anyone handling a client’s material, free no longer means private, and the choice of tier is now part of your compliance posture. Mechanics follow:

The mechanics: The prior Llama API terms stated flatly that “Meta will not use your Content to train any artificial intelligence models.” The new terms split that promise by tier. On the paid service, Section 5.1 states Meta acts as a processor and “will not use Content from Paid Services to train Meta Models.” This is a no training promise, not a no retention one: paid content can still be processed and retained for safety, security, abuse, and policy purposes. On the free service, Section 6.1 provides that “Meta may use your Content, including Inputs and Outputs, to train, develop, evaluate, and improve Meta’s artificial intelligence models, products, and services.” Before training on unpaid content, “Meta takes steps designed to disassociate that Content from your Accounts and your Meta Model API key; however, Meta may use Content for evaluation, safety, abuse, quality, and policy review without first applying those steps.” The routing rule is explicit in Section 6.2: “the Unpaid Services do not offer a mechanism to exclude specific traffic from training,” and “if you otherwise require that Content not be used for training, you must use the Paid Services for that traffic.”

Alongside the terms, Meta released a companion Acceptable Use Policy for the Model API. Two provisions are worth noting because they reach beyond generated content. It prohibits removing or altering “any watermark, label, metadata, or other provenance signal” on the model’s outputs, and it bars operating AI agents in ways that “evade human oversight or accountability,” including concealing an agent’s automated nature “where disclosure is required.” This is early, specific rule making aimed at agents and at the signals that tell people whether content was machine made. As an acceptable use policy, it binds Meta’s own API users, not third parties.

A note on reach, because it matters here: almost none of our readers use Meta’s first-party model API, which is still in limited preview. We lead with it anyway, because it is the plainest written proof of a direction every tool is drifting toward. Which brings us to the same move, seen from the other end.