Meta Will Cut Your AI Bill 92% If It Can Train on Your Prompts

Meta shipped Muse Spark 1.3 on September 2 with two price lists for the same model. The expensive one costs $1.25 per million input tokens. The cheap one costs $0.10. Tokens are the units a model bills by, so this is the line item your AI spend actually accrues in, and the gap between those two numbers is the whole story.

Meta’s own developer documentation explains the difference without any hedging. On the Standard tier, “your data is never used for training.” The Contributor tier “trades a lower price for permission to train on your prompts and completions.” OpenRouter’s listing for the Contributor model carries the same disclosure in one line: prompts and outputs may be used to improve Meta’s products.

Nobody is hiding anything here. That is what makes it worth writing about.

The size of the bribe

Tomasz Tunguz of Theory Ventures ran the numbers on September 3:

“In exchange for a 92% discount on input & a 95% discount on output, Meta retains the right to train.”

A 92% discount makes this a different product wearing the same name. At that spread, any team optimizing for cost lands on Contributor without a conversation, because the conversation only happens if somebody thinks to have it. Tunguz reads the design as familiar:

“Just like ads, the barter is simple : subsidized access in exchange for data.”

He’s right about the shape, and marketers should notice which side of the trade they’re on this time. When the barter was advertising, our industry bought the subsidized attention. Here we’re the thing being logged.

Why marketing is the exposed function

Two things are true about marketing teams that aren’t equally true elsewhere in a company.

The first is that marketing holds an unusual concentration of material that hasn’t shipped yet. Positioning that hasn’t launched. Pricing tests. Campaign copy under embargo. Customer research with names in it. The competitive teardown somebody pasted in for summarizing. All of it goes into a prompt window because that’s the fastest way to get work done, and the prompt window is what the Contributor tier is buying.

The second is that marketing usually picks its own tools. There’s rarely a procurement gate between a growth engineer and an API key, so the tier gets chosen by whoever’s moving fastest.

Cyberhaven’s 2026 report, built on data movements across 222 companies, found that 39.7% of all data movements into AI tools involve sensitive data, and that 32.3% of ChatGPT usage runs through personal accounts. Cyberhaven sells the software that catches this, so read the framing with that in mind. The direction still matches what anyone who has watched a marketing team work already knows. The report also predates Muse Spark by seven months, so it describes the habit, not this model.

Put the habit next to the price list and the outcome writes itself.

The governance answer nobody wants

Paul Roetzer, who founded and runs SmarterX and Marketing AI Institute, took a listener question on what happens to your data when an AI vendor is acquired or folds on his September 3 episode. His answer was that this isn’t yours to decide:

“you have to rely on your legal team”

He went on to say that a department leader doesn’t have the expertise for this call, and suggested using AI to compare a vendor’s terms against your own AI policy so an attorney sees the flagged parts before anyone signs.

That’s the correct answer and an uncomfortable one, because most marketing teams buying AI tools this year haven’t routed a single one through legal. A tier selection in a config file is a data-governance decision wearing the costume of a billing preference.

This is the third version of the same trade in six weeks

We wrote in August that Twitch switched on generative AI training for every channel by default and made creators opt out. We wrote last week that the Justice Department told a federal judge your published content is fair use as training data. Meta’s version is the most honest of the three and the most consequential: it isn’t taking your published work or flipping a default, it’s naming a price for your unpublished work and letting you accept.

Consent you’re paid for is still the strongest form of it. A 92% discount applied to a line item nobody reviews works more like a default with a rebate attached.

Five minutes of work, today

1. Find out which tier you’re on. In Meta’s API the model string is the tell: muse-spark-1.3 is Standard, muse-spark-1.3-contributor is the one that trains on you. Grep your codebase for the model name. This takes one search.
2. Ask the same question of every AI vendor in your stack, in writing. Does your data train their models by default, and is there a cheaper tier somebody may have selected? Discount tiers tied to training rights will not stay a Meta idea.
3. Write one line into your AI policy naming which tier is approved for client and customer data. You do not need a policy document to do this. You need one sentence somebody can point at.
4. Check whether your AI spend dropped recently without anyone explaining why. A sudden 90% fall in model costs has a cause, and this is now one of the likelier ones.
5. Send the terms to legal before the renewal, not after the incident. Roetzer’s answer is the right one, and it only works early.

Meta has priced something the industry has been taking for free, which is a kind of progress. The question for your team is whether anyone there knows what price got accepted, and when.