Meta Puts a $4.05 Price on Your Output Privacy • The Weighted Average

Meta is offering cheaper Muse Spark inference in exchange for permission to train on customers’ prompts and completions, according to September 3 reporting by TechCrunch. The choice puts a derived $4.05 per million output tokens on withholding that permission: subtract the $0.20 Contributor rate in Meta’s pricing documentation from the $4.25 standard rate confirmed in the contemporaneous report.

Reconstructed on September 7, 2026, from records available by September 3, 2026.

The discount purchases a second use of your work

This is not simply a cheaper model. It is a different transaction. Under standard pricing, Meta says prompts and completions are not used to train its models. Under Contributor pricing, the customer grants that permission. The token meter looks familiar; the rights attached to the tokens do not. A procurement comparison that erases that distinction compares unlike products.

The output saving is about 95.3%, calculated as $4.05 divided by $4.25. The input saving is smaller: the same sources show $1.25 per million uncached input tokens on standard pricing and $0.10 on Contributor, a derived 92% reduction. There is no single discount for every workload. Its effective size depends on how much input and output the application uses and whether its input can be cached.

Meta's training permission cuts output pricing by $4.05

Muse Spark output, US dollars per million tokens; different data rights

Meta pricing; TechCrunch · September 3, 2026

The graph is a comparison of commercial terms for model access, not an estimate of what a company’s confidential information is worth. A public documentation experiment and an internal acquisition memo may consume similar token counts while carrying utterly different disclosure consequences. The provider’s rate card cannot decide which one may be contributed. That decision belongs to the owner of the underlying information.

There is a second price hidden in the limits. Meta’s tier documentation gives standard teams 3,000 requests per minute and Contributor teams 100, a 30x difference. Token allowances are closer: four million versus three million per minute. The limits apply per team, not per key. Creating more keys therefore does not create an independent allowance for each application.

These are different bottlenecks. A workload made of many small requests can reach the Contributor request ceiling while leaving much of its token allowance unused. A smaller number of large jobs can behave differently. The practical task is to replay the application’s request distribution against both limits, not divide its monthly token bill by the advertised discount and declare the migration complete.

Caching complicates a simplistic comparison further. Meta’s prompt-caching documentation explains that matching prefixes can receive the cached-input rate. The relevant unit is therefore the actual mix of fresh input, cached input, and generated output in a request trace. An experiment that changes prompt structure at the same time as pricing tier may attribute an engineering improvement to the commercial discount, or miss the opposite effect.

The broader model push is visible in Meta’s Muse Spark 1.3 announcement, which frames the product around agentic and coding work. The argument here does not depend on a benchmark rank or any particular reasoning-mode availability. It depends on the more prosaic question of who may reuse the resulting work. The archive already treated frontier-model data custody as a purchasing feature; Meta now attaches an unusually explicit price to part of that bargain.

Separate shareable experiments from production memory

The obvious adopters are teams running experiments on public or deliberately shareable material. Contributor can be sensible when the application owner has affirmatively decided that its prompts and completions may become training data. That is narrower than deciding that a project is nonproduction. A test run can still contain customer records, internal specifications, or proprietary code copied into a debugging prompt.

An engineering team should define the eligible data class before selecting the cheaper model identifier. Keep the policy attached to the workload, document who approved the training permission, and prevent an experimental route from becoming the default for every application that shares a gateway. A routing rule is a commercial decision with data consequences, not just a latency optimization.

The cost of that separation is not disclosed by Meta’s price table. It includes data classification, integration review, and any additional operating path needed to keep contributed and noncontributed traffic distinct. The defensible financial statement is limited: the published output-rate saving contributes $4.05 toward those costs for every million output tokens processed. Whether that pays for the necessary controls depends on the organization’s measured workload and implementation cost. There is no evidence here for a universal break-even volume.

The strongest counterargument is that enterprises may classify too much material as unshareable and then pay for restrictions they do not need. TechCrunch’s reporting raises precisely that possibility: a visible discount can force buyers to distinguish genuinely proprietary work from routine experimentation. That would be a useful outcome. But classification should follow the data and contracts, not a target discount imposed on a team after the fact.

Standard pricing is also not a synonym for every privacy protection a buyer might require. A no-training statement does not by itself answer retention duration, regional processing, access controls, or deletion rights. Those questions still require their own documented answers. Treating a single permission flag as the entire governance agreement would make the apparent clarity of the price table misleading.

There is a useful cross-vendor check before granting those rights. Anthropic’s pricing documentation lists Sonnet 5 output at $10 per million tokens. Against Meta’s $4.25 standard rate, the difference is $5.75, or 57.5% less: ($10 − $4.25) ÷ $10. That is a comparison of raw output rates, not equivalent capability or total task cost. It shows that training permission is not a prerequisite for finding a lower sticker price; a no-training route still deserves a workload test.

The supply side has its own incentives. Broadcom’s larger AI revenue outlook shows suppliers planning for more infrastructure, while model vendors compete over the data that makes their services useful. Buyers should not expect all the value to appear as a lower price for equivalent rights. Sometimes the discount is payment in kind for giving the supplier something else.

The verdict is selective adoption, not reflexive rejection or automatic switching. Use Contributor where the data owner can approve the secondary use and the request limits fit the traffic. Keep sensitive workloads on an appropriately contracted route. Evidence that would change that judgment includes unsuitable training terms, failed segregation tests, excessive throttling, or controls whose cost exceeds the measured saving. Meta has made the trade visible. Visibility is an invitation to negotiate it, not permission to ignore it.