Muse’s Contributor Tier: Cheap Tokens, Real Trade-Offs

The Muse Spark contributor tier prices Meta’s newest model at $0.10 per million input tokens and $0.20 per million output tokens, against $1.25 and $4.25 on the standard tier — a 12.5× cut on input and a 21.25× cut on output for the identical checkpoint. Meta publishes no expiry date for it, and it is not a smaller model. It is a trade, and Meta’s documentation states the terms of that trade in plain language on the pricing page itself.

What you hand over is permission. The contributor SKU is described in Meta’s docs as heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models. The pricing table does not bury this in a footnote: the row for muse-spark-1.2-contributor is labelled “Used to improve our products,” and the row for muse-spark-1.2 is labelled “Not used to improve our products.” Two lines in the same table, one difference, a 12.5× price gap.

This guide takes the economics seriously rather than the headline. It covers exactly what the discount is worth per line item, the per-minute rate ceilings that quietly reshape which workloads can use it at all, the distribution catch that stops you routing to it through an aggregator, the parts of the bill the discount never touches, the near-identical precedent already running at Google, and a decision framework for which of your workloads can legitimately take the trade. If you want the product story instead, we covered the full Muse Code and Muse Spark 1.2 launch separately.

The cleanest way to understand the contributor tier is to notice what it is not. It is not a free tier with a usage cap. It is not a distilled or quantized variant. It is a second model ID — muse-spark-1.2-contributor — pointed at the same underlying checkpoint as muse-spark-1.2, carrying the same million-token context window and the same input modalities, priced an order of magnitude lower because of one clause about what happens to your traffic afterwards.

Meta’s own framing on the models documentation page is the marketing-side version of the same trade: the tier lowers the barrier to entry, gives you room to prototype, test integrations, and scale experiments without the usual cost overhead, in return for permission to train on your data. That is an accurate description of who it is for. The discount is aimed at the exploratory phase — the work where the tokens are numerous, the outputs are disposable, and the inputs are not somebody else’s confidential material.

Worth pausing on how unusually legible this is. Most data-use terms live in an appendix to a terms-of-service document that nobody reads at signup. Meta put the distinction in the pricing table, in the row header, next to the number — which means the trade is visible at the exact moment a developer is choosing a model string. That is a better disclosure pattern than the industry norm, and it is worth saying so even while arguing that plenty of teams should decline the offer. Compare it against the questions we suggest asking before you sign anything, in the procurement checklist for data-training clauses: this is the rare case where the answer is printed on the price list.

The three token line items do not discount uniformly, and the spread matters more than the headline. Dividing the Meta Model API list prices gives 12.5× on input ($1.25 against $0.10), 21.25× on output ($4.25 against $0.20), and 75× on cached input ($0.15 against $0.002).

Expressed the other way round, the contributor price is 8% of standard on input, roughly 4.7% on output, and about 1.3% on cached input. The chart below uses that framing because it makes the shape obvious: the more your workload leans on output tokens and cached context — which is to say, the more it looks like an agentic coding loop — the larger the proportional saving.

There is a tempting counter-argument that says heavy prompt caching on the standard tier gets you close enough to contributor pricing without giving up anything. The public data does not support it. At the time of writing, OpenRouter reports a 30-day weighted-average input cost of roughly $0.2967 per million tokens for the standard Muse Spark 1.2 listing. Back that rate out against Meta’s own list prices ($0.15 cached, $1.25 uncached) and it implies a cache-hit rate of about 86.7% across OpenRouter’s routed traffic. That figure is derived from the $0.2967 rather than an independent check on it, so treat it as a reading of the same number, not a second source for it.

The conclusion is the uncomfortable one for the caching argument. An exceptionally well-cached standard-tier workload still pays around $0.30 per million input tokens — roughly three times the contributor tier’s raw, uncached list price of $0.10. Caching does not close the gap; it narrows a 12.5× difference to about a 3× difference. And the contributor tier caches too, at $0.002, so a team that optimizes both tiers equally hard finds the proportional gap widening rather than closing.

This is the part that actually determines whether the discount is usable for a given workload. Meta’s pricing and rate-limits documentation states that rate limits control per-minute usage for your team and are set by pricing tier, then gives a two-row table: standard at 3,000 requests per minute and 4,000,000 tokens per minute, contributor at 60 requests per minute and 2,100,000 tokens per minute.

Read those two pairs against each other and the asymmetry jumps out. The contributor token ceiling is 52.5% of standard’s — a proportionate haircut you would barely notice on most workloads. The contributor request ceiling is 2% of standard’s. Those are not the same kind of constraint, and a team that plans capacity from the token number alone will be surprised.

The useful number nobody publishes is the crossover point — the average request size at which the token ceiling starts binding before the request ceiling does. On contributor it is 2,100,000 divided by 60, or 35,000 tokens per request. On standard it is 4,000,000 divided by 3,000, roughly 1,333 tokens per request. Below those thresholds the request counter is what stops you; above them, the token counter is.

That single division reframes the tier. Agentic coding loops are made of many small round-trips — a tool call, a file read, a short model turn, another tool call. Very few of those individual requests carry 35,000 tokens. So a contributor-tier team running the sort of workload Muse Code is built for will hit 60 requests per minute long before it comes anywhere near 2.1 million tokens per minute, and will be leaving most of its purchased token headroom unused at the moment it starts receiving 429s. The discount is real; the throughput is the thing you actually gave up.

Here is the finding that changes the practical calculus and that we have not seen anywhere else. The contributor tier is not listed on OpenRouter. We checked the obvious sibling URL to the standard model’s listing on August 7, 2026 and it returned a 404; only the standard meta/muse-spark-1.2 listing exists, at the $1.25 and $4.25 list rates, published there on August 5, 2026.

That has a straightforward consequence. You cannot get contributor pricing by editing a model string inside a tool that already routes through an aggregator. Taking the discount requires signing up directly with Meta, holding a team on the Meta Model API, and attaching a payment method — the docs note that billing triggers at a rising payment threshold or on the first of the month, whichever comes first, and that only a team admin can set the payment method. Whether that is deliberate design or simply a listing that has not landed yet, the effect today is the same: the cheap tier requires a deliberate, named relationship rather than a quiet configuration change.

Muse Code, the terminal agent, does not change any of this. Meta’s authentication documentation is clear that the CLI bills through the same Model API token metering and the same rate limits as direct API calls — there is no separate CLI quota system. Selecting the contributor model inside Muse Code inherits the 60 RPM ceiling exactly as an API integration would. If you are weighing the CLI itself against its rivals rather than weighing the pricing, that comparison lives in our Muse Code versus its terminal-agent rivals breakdown.

A per-token discount only discounts tokens. Everything else on the bill stays at full price, and because the token line has shrunk by an order of magnitude, the untouched items become proportionally enormous. Web-search grounding is the clearest example: Meta prices it at $2.50 per 1,000 queries, added on top of token costs, at the same rate on both tiers.

Run that against the token prices and the distortion is stark. On the standard tier, $2.50 of grounding costs the same as two million input tokens. On the contributor tier, the same $2.50 costs the same as twenty-five million input tokens. A search-grounded workload that was a rounding error against the token bill on standard can become the dominant line item on contributor — which means the effective blended discount on a grounding-heavy pipeline is far smaller than 12.5×, and you should model it that way before promising anyone a saving.

Two smaller mechanics are worth knowing, both in the developer’s favour. Meta’s injected system-prompt and scaffolding tokens are excluded from billing and from rate-limit counting, so only your own input and output tokens are metered. And there is no long-context premium on either tier: the per-token rate is the same whether the 1,048,576-token window is nearly empty or nearly full, which makes large-context experimentation unusually cheap on contributor specifically. For 429 handling, Meta documents the standard hygiene for both tiers — exponential backoff with jitter, starting around 500 milliseconds.

Reading the primary sources side by side, the contributor tier is closer to a rename than an invention. Google’s Gemini API Additional Terms of Service already draw the same line between a cheaper surface that trains and a paid surface that does not, and the wording is unambiguous in both directions.

For unpaid services, Google’s terms state that Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services. For paid services, the same document states that Google does not use your prompts — including associated system instructions, cached content, and files such as images, videos, or documents — or responses to improve its products. That is the identical trade Meta has priced. The difference is that Google draws the line at free versus paid, while Meta draws it between two paid tiers of the same model and attaches an explicit per-token number to it.

Naming a price is the genuinely new part, and it is worth sitting with. Meta has effectively published a market rate for training permission on your own traffic: on input tokens it is worth $1.15 per million to Meta, because that is the amount it is willing to forgo to get it. We have not seen a public number attached to it before. Our read is that this becomes the anchor others price against, and that within a year the interesting question in AI procurement stops being whether a vendor trains on your data and becomes what the vendor is willing to pay you not to.

The far end of this spectrum runs in the other direction: paying more for stricter handling rather than less for looser. Anthropic’s enterprise retention options are the clearest example, and we walked through what they change in practice in our piece on 30-day retention and zero-data-retention agreements. Between those two poles sits every other pricing shape currently on the market — including expiry-dated prepaid credit, which we covered when Qwen launched its prepaid token plan, and the older per-seat model documented in our guide to seat-based coding-tool economics. If you want the full taxonomy rather than this one shape in depth, how contributor pricing fits the four coding-tool pricing shapes is the companion piece.

The pricing analysis above is arithmetic. This part is judgment, and it is where most of the commentary stops short — every secondary write-up we found repeats the same four prices without asking who is actually allowed to use them. The framework below is ours, derived from Meta’s own tier definitions plus ordinary confidentiality and regulatory obligations. It is not vendor guidance, and it is not legal advice; treat it as the shape of the conversation to have with whoever owns your contracts.

The governing question is not “is our data sensitive.” It is “do we have the right to grant this permission.” A confidentiality clause in a client contract is not yours to trade away for a cheaper token rate, no matter how comfortable you personally feel about the risk. That distinction sorts most workloads in about thirty seconds.

Take a modest month of exploratory work: ten million input tokens and two million output tokens, none of it cached. On the standard tier that is $12.50 of input plus $8.50 of output, or $21.00. On the contributor tier it is $1.00 plus $0.40, or $1.40. The saving is $19.60, and the blended multiple is 15.0× — lower than the 21.25× output headline and higher than the 12.5× input headline, because the blend depends entirely on your input-to-output ratio.

Now add a thousand web-search grounding queries at $2.50 per thousand, charged identically on both tiers. Standard becomes $23.50 and contributor becomes $3.90. The blended multiple collapses from 15.0× to about 6.0×. Nothing about the token pricing changed; a single fixed-rate line item cut the effective discount by more than half. This is the arithmetic that turns an exciting headline into a planning number, and it is why the only honest way to evaluate the tier is against your own token mix rather than against the price list.

The volume point is worth stating plainly because it cuts against the tier’s own marketing. Because the discount is per token rather than a credit or a flat allowance, a small pilot saves a trivial amount of money in absolute terms — the month above saves less than twenty dollars. The teams that save materially are the ones running genuinely high token volumes, and those are disproportionately the teams whose traffic is worth the most to a model developer. The trade is priced accordingly. That is not a criticism; it is the reason to check whether you are the party getting the better end of it.

One availability caveat before you plan around access. At launch the tier was reported as limited to selected countries. Meta’s Model API product page now describes the product as being in public preview with expanded global access, and no current page we checked names which countries were ever excluded — so treat the earlier restriction as launch-era information that may no longer apply, and verify availability for your own jurisdiction against Meta’s live documentation rather than against any secondary summary, this one included.

Looking forward, the most likely outcome is that this pricing shape spreads rather than stays a Meta curiosity. Named, priced data tiering solves a real problem for developers who genuinely do not care about training rights on throwaway work, and it gives model developers a cleaner acquisition channel than scraping. Expect other vendors to follow with their own contributor equivalents, expect procurement teams to start asking which tier a supplier’s agency or contractor is actually using on their account, and expect the answer to that question to become a standard line in AI vendor questionnaires well before the end of 2027. Getting your own answer straight now is cheaper than retrofitting it later — which is exactly the work our AI transformation engagements start with.