Meta's Cheaper Tier Paid for With Data Now Covers Muse Spark 1.3 - Prices and Rate Limits Unchanged Since August

The Contributor tier we covered in August now also lists the latest muse-spark-1.3, while its per-token prices and rate limits are unchanged. A second look at the gap against standard pricing, the Standard tier's explicit "not used to train" line, and the drop in rate limits.

Meta API pricing data-privacy LLM coding

The two-tier structure on Meta’s developer documentation page “Pricing and rate limits” is something this site already covered in August, in its article on the beta release of Muse Code and Muse Spark 1.2. Checking the pricing page again now, the latest generation muse-spark-1.3 and muse-spark-1.3-contributor have been added to the two tiers, while the per-token prices and rate limits are unchanged from what we recorded in August¹. With TechCrunch picking the arrangement up², it is worth going back over the terms.

One tier is the Standard tier; the other is the Contributor tier, described this way¹:

Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models.

The gap is not small. Putting the two tables on that page side by side¹:

Per 1M tokens	Standard tier	Contributor tier
Cached input	$0.15	$0.002
Input	$1.25	$0.10
Output	$4.25	$0.20

Worked out, that is 92% off input, about 95% off output, and about 99% off cached input. Those three figures match what we recorded in August.

The list of models covered has grown. The current page shows muse-spark-1.3, muse-spark-1.2 and muse-spark-1.1 on the Standard side, and muse-spark-1.3-contributor and muse-spark-1.2-contributor on the Contributor side¹. In August, Standard listed only 1.1 and 1.2, and Contributor only 1.2-contributor. So the latest 1.3 generation now comes with the same data-for-price option.

The line that pairs with it sits on the Standard tier

It is easy to miss, but the Contributor tier is not the only one on this page that spells out a condition. The Standard tier’s description reads¹:

Standard pricing; your prompts and completions are not used to train Meta models.

So Meta has placed, on the same page, a price for not being trained on and a cheaper price for allowing it. How data is handled is not buried in terms of service here; it is a line item in the price list.

That is where the practical point sits. Compare APIs on unit price alone and the Contributor tier’s $0.10 reads as “Muse Spark is cheap.” The cost of that cheapness is printed just under the heading of the same table. When a company evaluates models and the person building the price comparison is not the person reading the terms, this is exactly the kind of item that falls out of the table.

Meta also states the use it has in mind, in the second half of the description: lowering the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable¹.

The rate limits are not built for production

The rate limit table makes the same point more concretely¹:

Tier	Requests per minute (RPM)	Tokens per minute (TPM)
Standard	3,000	4,000,000
Contributor	100	3,000,000

RPM is one-thirtieth of Standard, while TPM is down 25%. These figures are unchanged from August as well. Total token throughput is not squeezed nearly as hard as request count. A production workload that fires many short requests at high frequency will not fit through this tier; an experiment running a small number of long requests will.

Limits also apply per team, not per API key¹ — so one person’s usage counts against everyone else’s. Muse Image (muse-image-1.0) has its own separate limit of 150 requests per minute¹.

The overall billing model is described as pay-only-for-what-you-use: text models per token, image generation per image, and Muse Voice Transcribe per minute of audio processed, with no minimums or upfront commitment¹.

What belongs in a comparison table of “cheap models”

Meta launched the Meta Model API alongside Muse Spark 1.1 in July, and in August released the terminal coding agent Muse Code in beta alongside Muse Spark 1.2. This tier has widened along with that coding-focused expansion.

That price moves model selection is visible in the data. An analysis of Ramp’s spending data showed that companies are not paying for top-tier models as much as one might expect. Cheap tiers work — which is why it matters whether what the cheapness costs is printed alongside it.

Price cuts elsewhere continue. When OpenAI cut GPT-5.6 Sol by 20% on input and 33% on output in August, the API changelog noted those were promotional rates available at least through November 21, 2026. What “cheap” consists of differs by provider.

The pricing page carries no last-updated date or changelog, so what changed and when cannot be traced from the page alone. TechCrunch has reported on the arrangement, and says Meta did not respond to its question about the new pricing model². The page’s own wording is the only official explanation available right now.

If your organization is deciding whether to use this API, the order of checks is clear enough: before looking at the unit price, look at whether the model name ends in -contributor. If it does, you are using it on the understanding that the code and design fragments you put in those requests become training material for a future model.