Meta’s documentation for its coding agent has a table headed “Models and pricing”. It has two rows.
muse-spark-1.2-contributor Used to improve our products. 1M $0.10 $0.002 $0.20
muse-spark-1.2 Not used to improve our products. 1M $1.25 $0.15 $4.25Columns: context window, input per million tokens, cached input, output. Read across, and the cheaper row costs a fraction of the other one, with a sentence in the middle explaining why.
The thing most coverage gets structurally wrong
This is usually described as a discount tier you can opt into, and that framing shapes every question people ask about it. Is it the default? Can you switch? Where is the setting?
The documentation shows something else, and Meta’s own heading says it: the table is titled “Models and pricing”. These are two model identifiers. muse-spark-1.2-contributor and muse-spark-1.2 are different strings in an API call, the way you would choose between two model sizes.
That moves the question rather than answering it. There is no account setting and no consent screen where a tier gets selected. There is a model name in a configuration file, and the question becomes which string that file contains. For anyone writing the call themselves, that is now trivially checkable. For anyone who installed the CLI with the one-line script Meta puts on the same page, it is not, because what the installer writes into that file is not documented.
What the difference costs
Three factors, taken from the two rows:
Input: 12.5 times. $1.25 against $0.10 per million tokens.
Output: 21.25 times. $4.25 against $0.20.
Cached input: 75 times. $0.15 against $0.002. This is the largest gap in the table by a wide margin, and it is the one nobody quotes. For an agent that re-reads the same repository across a long session, cached input is not a rounding error. It is much of the bill.
The context window is identical at one million tokens. Whatever is being traded here, capability is not visibly part of it.
The exact wording, and why it is wider than reported
The rows do not say “used for training”. They say:
“Used to improve our products.”
And, for the expensive model:
“Not used to improve our products.”
Reporting on this consistently renders it as training data. MacRumors describes Meta “offering a discount for users who give the company permission to use their prompts and completions to train future AI models”, and that is a fair summary of the trade. It is also narrower than what Meta’s own table says.
“Improve our products” covers training. It also covers evaluation, safety review, product analytics, and whatever else the phrase reaches. A company that meant only training could have written only training, and this one did not: the words “training” and “train” appear zero times on the page, counted in the rendered text. If you are assessing what leaves your machine, assess the sentence Meta wrote, not the one that got reported.
What the documentation does not say
It does not say which model is the default. Several secondary sources state that Muse Code defaults to the contributor model after installation, which would mean the data sharing is the starting state rather than a choice. Search results attribute this to “early user reports”. Meta’s own product page contains no statement about a default, and the phrase “by default” appears there once, describing multi-agent behaviour and nothing to do with pricing.
If that claim is true it is the most important fact about this product, and it is currently sourced to users rather than to the vendor.
It does not state rate limits. The page has no rate-limit figures at all. Figures do circulate, putting the cheap model at 60 requests per minute against 3,000 for the standard one, which would mean the discount is paid for twice, once in data and once in throughput. That comes from reporting, not from Meta’s page.
The practical question, for once, is easy to check
Open the configuration your agent uses and look at the model string. If it ends in -contributor, the row that applies to you says your prompts and completions are used to improve Meta’s products.
That is a five-second check on a machine that already has it installed, and it beats any claim about defaults, including the ones repeated in this article. The setting is a name in a file, not a switch buried in an account page. What nobody outside Meta can tell you is what that name will be before you run the installer.
One aside for anyone verifying where any of this comes from. The documentation lives on developer.meta.com and the install script on dev.meta.ai, two different domains. Both carry organisation-validated certificates issued to Meta Platforms, Inc., so both are genuine — checked, no finding. It is worth naming anyway, because a split like this is the exact situation in which something like dev-meta.ai would go unnoticed, and the second domain is the one you pipe into a shell.
Two things worth thinking about before changing the model string. The saving is real and large, particularly on cached input, and for hobby projects or open-source work it may be exactly the right trade. And for a team pushing proprietary code through an agent, the sentence in that row is the entire question, regardless of the price.
Sources
- Meta, “Muse Code” product page: https://developer.meta.com/ai/products/muse-code/ (retrieved 28 August 2026). Source of the two model identifiers, both quoted product statements, the context window, and all six prices. The page is rendered client-side and returns only a navigation shell to a plain HTTP fetch; the content above was read from the fully rendered page. The domain carries an organisation-validated certificate issued to Meta Platforms, Inc., checked on 28 August 2026. The page states no default model and no rate limits; both absences were counted in the rendered text, not inferred.
- MacRumors, “Meta Muse Code for Mac”, 5 August 2026: https://www.macrumors.com/2026/08/05/meta-muse-code-for-mac/ (retrieved 26 August 2026). Source of the quoted characterisation of the discount as permission to use prompts and completions for training, and of the circulating rate-limit figures.
- The Register’s coverage of the launch lists the standard prices and the sentence “Discounts are available”, without the condition attached to the discount. It is cited here only to note what it does not contain, because it is frequently given as the source for the trade.
