If you read one line: Meta's Muse Spark 1.2 charges two prices for the same model — a standard tier, and a "contributor" tier that's roughly 8–10× cheaper because your prompts and the model's code become Meta's training data. The discount is real. Whether to take it isn't an ethics question — it's a data-classification question, and it has a clean answer per repo.
The math, first#
Same model, two rates (Simon Willison):
| Standard | Contributor | |
|---|---|---|
| Input / 1M | $1.25 | ~$0.10 |
| Output / 1M | $4.25 | ~$0.20 |
That's about 12× cheaper on input and 21× cheaper on output. A coding agent reads far more than it writes — big context windows, whole files pulled in, small diffs out — so the blended saving on a realistic workload lands around 85–90% off your model bill. If you're spending $2,000/month on Muse Spark tokens, the contributor tier is a ~$200–300 line item for identical model quality.
There's one structural limit: the contributor tier is capped at about 60 requests per minute (MarkTechPost). That's plenty for a human at a terminal and too little for high-throughput batch pipelines — so it's aimed at interactive coding, not a data-labeling farm.
(Pricing is from launch-week reporting, not a first-party Meta page — confirm the exact rate before you budget.)
What you're actually paying with#
The discount isn't a promotion — it's a purchase. In exchange for the lower rate, Meta trains future models on your prompts and on the completions it sends back. That means both halves of the exchange leave your control: the code and instructions you send in, and the code the model writes out.
For a lot of work, that's a shrug. For some work, it's the whole ballgame. The mistake is treating it as one policy for your entire org instead of a per-workload call.
The five-question test#
Before you route a repo through the contributor tier, ask:
- Would I publish this code openly? If it's already destined for a permissive-licensed public repo, training on it costs you nothing you were keeping.
- Is there a secret in the context window? API keys, tokens,
.envcontents, connection strings — if any of that can land in a prompt, the discount is off the table until you strip it. (See redacting secrets before they reach a model.) - Is this someone else's code or data? Client work under an NDA or DPA, or anything with personal/regulated data, almost always forbids sending it to a train-on-your-data tier. That's a contract question, not a price question.
- Is the structure itself the moat? For some products the proprietary business logic or architecture is the defensible asset. Feeding it to a competitor's training run is a strategic cost no per-token discount offsets.
- Can I isolate the workload? If you can cleanly separate throwaway/OSS work from sensitive work, you can take the discount on the former without exposing the latter.
If the answers are yes, no, no, no, yes — take the discount. If any of the middle three flips, stay on the standard (no-train) tier for that repo.
How to actually run both#
You don't have to pick one tier for everything. Because the Muse Model API is OpenAI- and Anthropic-compatible, the clean pattern is route by sensitivity:
- Contributor tier for scratch projects, prototypes, learning, and open-source contributions — the 85–90% saving with nothing to lose.
- Standard (no-train) tier for the company's real codebase, client work, and anything touching secrets or regulated data.
Wire the split at your proxy or router keyed on the repo, not on a per-developer habit — humans forget which tier they're on, and the whole point is that the mistake is expensive in exactly one direction. Then read the license and data-processing terms before the first real request; a discount you can't legally use isn't a discount.
The bigger pattern#
Meta's contributor tier is the loudest version of a lever that's spreading across model vendors: cheaper tokens in exchange for training rights, usually offered opposite a paid "no-train" default. Expect more of it, and expect the discounts to get steeper as labs compete for fresh, real-world code to train on. The durable skill isn't picking the right vendor this week — it's having a standing data-classification rule so that when the next 10×-off tier lands, you already know which of your workloads can take it and which can't. The teams that get burned won't be the ones who used the cheap tier. They'll be the ones who used it without deciding which code was on it.
