Meta’s cheap Muse Spark tier trains on your prompts and code

What we checked

• What do Meta’s own Muse Code pricing page and terms say about contributor-tier data use, retention, and rate limits? confirmed. Official docs price Contributor Muse Spark at $0.10/$0.002/$0.20 per million input/cached/output tokens versus $1.25/$0.15/$4.25 Standard, in exchange for using prompts, code, files, context, and responses to train future models. Contributor API IDs are limited to 150 requests per minute and 3 million tokens per minute versus 3,000 and 4 million on Standard. Terms Section 6 lets Meta retain that content for training, law, and policy review after steps designed to disassociate it from the account. [1][3][7][6]
• Was the Muse Code launch already covered separately in prior publication surfaces not visible in this basket? confirmed. Meta published a developer blog and a research blog on August 5, 2026 announcing Muse Code beta and Muse Spark 1.2, including contributor versus standard model IDs and a sales path for zero data retention. CNBC covered the same-day launch and quoted AI chief Alexandr Wang on a contributor tier more than ten times cheaper than pay-as-you-go. A later Meta post on August 31, 2026 took Muse Code out of beta and began rolling out subscriptions. [9][14][12][15]
• Are Contributor-mode token prices uniformly about 20 times cheaper than Standard? partly confirmed. The 20x headline matches output tokens ($4.25 versus $0.20, a 21.25x cut) and is close to the 20x gap in requests per minute (3,000 versus 150). Input is 12.5 times cheaper ($1.25 versus $0.10); cached input is 75 times cheaper ($0.15 versus $0.002). Wang described the tier as more than ten times cheaper, which matches the documented input discount rather than a uniform 20x. [1][12][2]
• Do Muse Code subscriptions default to Contributor mode, and is that mode throttled well below Standard? partly confirmed. Subscriptions default to Contributor; a new computer turns it on and shows a notice, and Terms Section 3.3 treats Coding Harness Subscriptions as Contributor unless the setting is changed. Turning it off reduces plan usage without changing the plan. API Contributor RPM is 20 times lower, but TPM is only about 25 percent lower. Headless runs with no settings file default Contributor off. [6][7][3][1]

What we found

• Meta’s pay-as-you-go table prices Muse Spark Standard at $1.25 per million input tokens, $0.15 cached, and $4.25 output. Contributor variants muse-spark-1.2-contributor and muse-spark-1.3-contributor are $0.10, $0.002, and $0.20. Muse Wonder has no Contributor prices on the API. OpenRouter lists the same Contributor rates and states that prompts and outputs may be used to improve Meta’s products. Cached-input discounts are large enough that real bills can fall far below the headline input rate. [1][2]
• The Meta Model API Terms, last updated October 9, 2026, split Standard (Section 5) from Contributor Services (Section 6). On Standard, Meta will not use request content to train models. On Contributor, Meta may use Content to train, develop, evaluate, and improve models, products, and services, after steps designed to disassociate it from accounts and API keys; evaluation, safety, abuse, quality, and policy review can skip those steps. Content previously sent on Standard becomes Contributor content if later included as context. You must not submit sensitive, confidential, or personal information, including software code you are required to keep confidential. Meta’s IP indemnity in Section 17.2 applies to Standard Services, not Contributor. [3][4][5]
• Muse Code subscriptions default to Contributor. Help text says you get more usage in exchange for letting Meta use prompts, code, files, context, and responses to train future models. The setting is per computer; the first interactive session on a new machine turns Contributor on and shows a notice. Headless muse exec with no settings file has it off. Per-run flags and MUSE_CONTRIBUTOR_MODE exist; organizations can lock it off via managed configuration. Turning it off is prospective only: Section 6 rights survive for content already submitted. Everyday Usage is $5 per month for 10–50 prompts every 5 hours; High, Power, and Maximum are $15, $50, and $200. [6][7][8][3]
• On the API, Contributor Spark is limited to 150 requests per minute and 3 million tokens per minute, versus 3,000 and 4 million for Standard Spark. The August 5 launch post described Contributor as rate-limited by tokens in a rolling 5-hour window, not by request count, and listed Standard pay-as-you-go at $0.15/$1.25/$4.25 per million cached/input/output. The cheap path is throttled on requests much more than on tokens. Muse Spark is a reasoning model; thinking tokens are billed as output. [1][9]
• Zero Data Retention is not self-serve. Meta enables it at account level after a sales request, with no obligation to grant it. When on, content is processed in real time and not stored at rest after the response, with listed exceptions. ZDR Terms Section 2.5 excludes Contributor Services. Help pages list Contributor mode, file uploads, background requests, web search, and stateful storage as unavailable. ZDR does not cover third-party platforms, gateways, or hosted routing. Wang told CNBC Meta was starting to accept ZDR requests as an enterprise feature. [10][11][12][9]
• Wang told CNBC Meta is competing on price, not capability, and that Contributor is more than ten times cheaper if developers opt in to help improve the model. Meta’s current Terms instead treat Coding Harness Subscriptions as Contributor by default. Q2 2026 Exhibit 99.1 shows advertising revenue of $59.363 billion and total revenue of $60.801 billion. Meta has not said Contributor prompts are used for advertising; the documented use is training and related evaluation. [12][3][13]

What it means

Treat the model ID and the data election as separate controls. For pay-as-you-go, call muse-spark-1.x when the repo or customer contract cannot go into a training set; use -contributor only on public or throwaway work. For Muse Code, do not rely on Wang’s opt-in phrasing: pin Contributor off in /settings on every machine, set --contributor-mode off or MUSE_CONTRIBUTOR_MODE=0 in CI, and prefer org-managed configuration so a new laptop cannot flip it back on. Do not put personal data, secrets, or NDA code on Contributor; the Terms say not to, and Standard IP indemnity does not apply. If you need ZDR, ask sales, budget Standard prices, and call Meta’s API directly rather than a gateway. The 12.5x input and 21x output discounts are real; so is the loss of training-use protection and most of the request-rate budget.

Still open

• How long disassociated Contributor content is kept in training corpora, and whether account deletion removes it from those corpora.
• Why the ZDR help page says policy-violation retention can last up to two years while the ZDR Terms say flagged content may be kept up to thirty days.
• How Exhibit C (Meta says it does not receive hosted-platform content) interacts with OpenRouter forwarding Contributor requests to Meta.
• The exact usage multiplier when a Muse Code subscription turns Contributor off.