Quick answer. Muse Code's contributor tier costs $0.10 per million input tokens and $0.20 output — 12.5x and 21.25x cheaper than standard — because Meta trains on your prompts and completions. It is selected by a model id, not a signed agreement, and it is throttled to 100 requests per minute. Never point it at proprietary code.
When Meta launched Muse Code on August 5, 2026, the number everyone repeated was the discount: a contributor tier "more than 10 times cheaper" than pay-as-you-go, in the words of Meta's Chief AI Officer Alexandr Wang.
The actual mechanism deserves more attention than the number. Meta did not price a discount tier. It priced a transaction — and it is the first major vendor to put an explicit, public price on the thing every AI company wants and cannot easily buy.
What exactly is the contributor tier?
Two model ids, two prices, one difference:
| Standard | Contributor | Delta | |
|---|---|---|---|
| Input / 1M tokens | $1.25 | $0.10 | 12.5x |
| Output / 1M tokens | $4.25 | $0.20 | 21.25x |
| Cached input / 1M | $0.15 | $0.002 | 75x |
| Rate limit | 3,000 RPM | 100 RPM | 30x fewer |
| Meta trains on your data | No | Yes | — |
The clearest public statement of the trade, as restated in third-party model catalogues, is blunt: your inputs and outputs are used to train and improve Meta's AI models.
Wang's "more than 10 times cheaper" framing, if anything, undersells it. On output tokens it is over 21x, and on cached input it approaches 75x.
Why is Meta doing this?
Because agentic coding traces are the scarcest training data in the industry right now, and there is no good way to buy them.
Public code on GitHub is finished — everyone has scraped it. What nobody has at scale is the trajectory: the prompt, the plan, the tool calls, the failed attempt, the correction, the test run, the fix. That sequence is what teaches a model to act rather than autocomplete, and it only exists when a real developer solves a real problem with an agent watching.
Forbes put the strategic logic well — the tier turns compute into training data Meta would otherwise have to license, scrape or manufacture.
The financial context sharpens it. Meta's Q2 2026 results, reported a week before the Muse Code launch, showed free cash flow collapsing to $784 million from $8.55 billion a year earlier, with capital expenditure at $31.1 billion and full-year guidance raised to $130–145 billion. Against that backdrop, subsidising inference to acquire proprietary training data is a rational trade — and Mark Zuckerberg's stated rationale, that there is significantly higher margin in selling intelligence than selling compute, is the whole thesis in one sentence.
Meta is not being uniquely predatory here. Google's free Gemini tier trains on input with human review. Most free AI tiers have always been paid for with data. Meta's difference is that it put a price tag on it and made it a product decision instead of a buried clause.
What is actually risky about it?
Not the trade itself. The trade is honest. The risk is how the choice gets made.
The tier is selected by a model id — a single string in a config file or an environment variable. There is no separate contract, no signed data-sharing agreement, no procurement review, no admin console toggle that a security team can lock.
That means a 12–21x line-item saving sits behind a change any engineer can make in thirty seconds. As one commentator put it, the cheap tier will end up in a lot of .env files by Friday, chosen by whoever was told to keep the AI bill down.
This is the real governance failure, and it has nothing to do with Meta's intentions. It is a design in which the cheapest default is also the one that gives away your source code, and the decision is made at the wrong altitude — by an individual optimising a bill, not by anyone accountable for IP.
It uploads your CLAUDE.md to Meta by default
This one deserves its own heading, because it is easy to miss and it compounds everything above.
Muse Code loads your machine-wide agent instruction files — ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md — from outside the workspace, and places their full contents in the first request it sends. Meta's own configuration documentation confirms the behaviour: your machine-wide user rules always load. There is an opt-out flag, --no-foreign-personal-context, but no interactive consent prompt.
On the standard tier that is a privacy footnote. On the contributor tier it is not: those files are prompt content, and prompt content is exactly what the contributor tier grants Meta training rights over. If your personal CLAUDE.md contains internal architecture notes, naming conventions, credentials guidance, client names or unreleased product context — and for most working engineers it contains at least one of those — that material is in scope from your very first request.
Two practical takeaways. Audit what is actually in your machine-wide instruction files before running any agent that reads them, not just this one. And if you use the contributor tier at all, run it with --no-foreign-personal-context as a default rather than an afterthought.
Why is this different from other AI data mistakes?
Because it is the one you cannot undo.
Nearly every AI data incident is remediable. Logs get deleted. Retention windows expire. Vendors sign amended agreements. Access gets revoked.
Training is different. Once your code is absorbed into model weights, there is no extraction procedure. Under GDPR, an Article 17 erasure request is effectively unsatisfiable against a trained model — you cannot delete a gradient. The same applies to trade-secret status: code that has been used to train a third party's commercial model is difficult to characterise as having been kept secret.
For most organisations this is the whole argument. Not "Meta might misuse it" — simply that the action is irreversible, and irreversible decisions should not be reachable by config string.
Is the discount even real at scale?
Less than it looks, for three reasons.
The rate limit is the binding constraint. 100 requests per minute against standard's 3,000 is a 30x throughput cut — applied to a product whose headline feature is parallel subagents running across git worktrees. You cannot run Muse Code's flagship capability at contributor speed. Almost nobody reported this alongside the price.
Cheap tokens are not cheap outcomes. The meaningful metric is cost per solved task, not cost per token. A weaker model burns more turns, more retries and more context reaching the same result. On measured cost-per-solved-task, the gap against frontier models compresses to low single digits — not the order of magnitude the token rates imply.
Free credits carried the same terms — added after the fact. There is a first-hand, on-the-record report that Meta attached small print stating that content may be used for product improvement to credits that had already been granted at the earlier Muse Spark 1.1 launch, under terms that were not present when users received them. If that reading holds, the evaluation on-ramp was also a data transaction regardless of which model id anyone selected — which undercuts the "explicit, transparent trade" framing Meta otherwise earned credit for.
This has gone essentially unreported, and it is arguably a sharper story than the default-upload behaviour, because it means the terms you evaluate under are not necessarily the terms you end up operating under. It is also the strongest argument in this entire piece for treating the tier as a policy decision rather than a configuration one: a control you can only exercise at signup is not a control at all.
The tier nobody wrote about is the interesting one
Lost in the noise about the 21x discount: Muse Code's standard tier is aggressive on its own.
At $1.25 input and $4.25 output per million tokens, it undercuts Claude Opus 5 ($5/$25) by roughly 5.9x on output and GPT-5.6 Sol by a similar margin — with no training rights, no throughput penalty, and the full 3,000 RPM.
For most engineering organisations, that is the offer worth evaluating. It is a real price cut on frontier-adjacent capability with no strings, and it went almost entirely unreported because a 5.9x discount is a less exciting headline than a 21x one.
What should teams actually do?
Decide the tier at the policy layer, not the config layer. Pin the model id in shared configuration, and treat a change to it as a reviewable event. If your platform team manages agent configuration centrally, this is a two-line policy. If every developer configures their own, you have a problem that predates Muse Code.
Block contributor egress for repositories that matter. Where you can enforce it at the network or proxy layer, do. Convention is not a control.
Use contributor deliberately, not accidentally. It is genuinely excellent for open-source work, throwaway prototypes, public documentation, scratch repositories and learning. Training on code that is already public costs you nothing. That is a real and legitimate use of a very cheap tier.
Default to standard for everything else. The no-strings price is already good. Paying 12x more to keep your source code out of a competitor's training set is not a difficult business case to make — it is rounding error against one engineer-hour.
Ask the questions Meta has not answered publicly. As of early August 2026 there is no published retention window, no stated processing region, no SSO or admin console, no server-side audit log, no DPA and no scoped compliance attestation for the Model API. Meta holds SOC 2 and ISO 27001 certifications for other business products; do not assume they extend here. For a regulated environment, absence of published terms is itself a blocker.
The bottom line
The contributor tier is not a trap. It is a clearly priced trade, and Meta deserves some credit for making it explicit rather than burying it in a terms-of-service update.
It is also the first time the price of your source code has been printed on a public pricing page — roughly a 12 to 21 times discount on inference. That is a genuinely useful number to know, and worth deciding about deliberately rather than discovering in a billing report.
Use it for open code. Keep it away from everything else. And make that decision somewhere more durable than an environment variable.
For the full product picture, see our Muse Code guide and how it compares to Claude Code.
FAQ
What is the Muse Code contributor tier?
A discounted pricing tier at $0.10 per million input tokens and $0.20 output, versus $1.25/$4.25 on standard. In exchange, Meta uses your prompts and completions to train its models. It is selected by using a different model id.
Does Meta train on my code with Muse Code?
On the standard tier, no. On the contributor tier, yes — that is what the discount buys. There are also reports that free credits carry product-improvement terms, which is worth verifying against current terms before evaluating.
How much cheaper is the contributor tier?
12.5x on input, 21.25x on output, and around 75x on cached input. But it is rate-limited to 100 requests per minute versus 3,000, so effective throughput is 30x lower.
Can I use the contributor tier for private repositories?
You can technically, but you should not. Once code enters model weights it cannot be extracted or deleted, which makes GDPR erasure requests unsatisfiable and complicates trade-secret protection.
How do I stop my team using the contributor tier by accident?
Pin the model id in shared configuration and treat changes as reviewable. Where possible, block contributor endpoints at the network or proxy layer. Because the tier is one config string, convention alone is not a control.
Is the contributor tier good for anything?
Yes — open-source work, public documentation, prototypes, scratch repositories and learning. If the code is already public, training on it costs you nothing and the economics are excellent.
Does Muse Code offer enterprise data protections?
As of early August 2026 Meta has not published a retention window, processing regions, SSO, an admin console, server-side audit logs or a DPA for the Model API. Existing Meta compliance certifications cover other products and should not be assumed to extend here.
Is Meta's approach worse than other AI vendors?
Not fundamentally. Google's free Gemini tier trains on input with human review, and most free AI tiers are paid for with data. Meta's difference is that it priced the trade openly instead of burying it — which is arguably more honest, but also makes the cheap option easy to select without review.