Meta will cut your API bill 22x for permission to train on your prompts. The rate that buys is $0.10 in and $0.20 out, which is precisely what Poolside's Laguna S 2.1 already sells for with nothing attached.

Before the numbers: Meta has not published any of them

This normally goes at the bottom. It goes here because of how far it reaches into everything below. Meta's announcement post for Muse Code and Muse Spark 1.2 contains no dollar figure anywhere. Not the standard rate, not the contributor rate, not the rate limits, not the context window, and no statement of what the contributor tier does with your data. The developer product page for Muse Code carries no pricing either.

To be fair where fairness is due, this is not a company that documents nothing. The announcement links a three-page evaluation methodology report that names every comparator, says which harness each model ran under, specifies pass@1 across five attempts, and volunteers that the setup "may not be specifically tuned for proprietary third-party models." That is a better benchmark disclosure than most launches manage. It contains no prices, no rate limits, no context window and no data policy. Meta documented the part it was measured on and left the part you get billed on to the press.

So the numbers below come from the launch email and the Model API rate card as reproduced by outlets that received them. Five sources agree to the cent on all six price points, which is strong corroboration and we are confident quoting it. Be aware the corroboration is thinner past that: the rate limits and the context window essentially rest on one detailed write-up, and another source we cite states flatly that Meta disclosed no limits or context at launch. One of the five says it verified against Meta's Model API documentation behind the console, so a first-party page may exist; it is not one you can reach without an account.

There is a specific reason to want a public page here. A tier whose entire product is a data-use permission is a tier where the exact wording of the permission is the thing you are agreeing to, and that wording is not published anywhere we could find. One outlet notes Meta has not clarified whether the contributor grant is training only, or also covers evaluation, red-teaming and product analytics. We could not resolve it either.

Two things we can check against our own records. Meta's standard tier is unchanged from Muse Spark 1.1, which we logged at $1.25 and $4.25 back in July, and it still is, so the entire pricing story of this release is the new SKU. And our note on that July entry reads "pricing via Meta API console reported by Reuters and Bloomberg, not on Meta's own docs page at launch." This is the second Meta model API launch in a row where the rate card reached the press before it reached a page Meta owns. At some point that stops being an oversight and starts being the process.

One rate card, three different discounts

The two tiers are not one discount applied to a rate card. They are three discounts of very different sizes, and the spread between them is the most informative thing on the page.

Look at the cached line on its own terms rather than as a multiple. On the standard tier, a cached token costs 12% of a fresh one, which is an ordinary industry ratio. On the contributor tier it costs 2%. Meta did not apply its discount evenly and then round; it went six times deeper on cache than it did on input.

Cached input, in a coding agent, is your repository. It is the hundreds of thousands of tokens of source that get resent on every turn of a long task and hit the prefix cache. That is the leg Meta discounted hardest, and it is also the leg that carries the most of your actual code per dollar forgone. Run a million tokens at a 90% cache hit rate and the standard tier bills $0.26 against just over a cent on the contributor tier, so the effective saving climbs from 92% to about 95.5%. Pricing shaped like that pulls in whole repositories rather than clever prompts.

What Meta is bidding for a million tokens of your code

Discount multiples are the wrong unit for this decision. You are not buying inference at a discount, you are selling something, and the question a seller asks is the price. Subtract one rate card from the other and you get it directly: Meta is offering $1.15 per million uncached input tokens, $0.148 per million cached input tokens and $4.05 per million output tokens.

Those are per-meter bids. What a month is worth depends on your mix, so here are four profiles: three intensities of agentic coding, then one deliberately different shape to show what moves the answer. To be clear about what these are, the token volumes are our assumptions, not measurements of Muse Code, which has been public for a day. The rates applied to them are the published ones.

Divide the bid by the traffic and the first three rows converge: the two larger ones both land on $0.347 per million tokens, and the hobbyist is only fractionally above at $0.382. Scale does not move the price. Across any mix roughly shaped like agentic coding, Meta's offer is about thirty-five cents a million tokens whether you are one person or a department.

Shape moves it enormously, though, which is why the last row is there. Drop the cache and weight the traffic toward output and the same arithmetic gives $3.08 per million, roughly nine times the coding bid, because output is where the standard card is expensive. The practical reading: Meta pays best for traffic that is mostly the model talking, and worst for traffic that is mostly your repository sitting in a cache. Given that the cache discount is the deepest one on the card, those two facts sit oddly together. The cheapest thing to send is the thing Meta pays least for, which suggests the cache rate is about getting the repository in the door rather than about what the tokens are worth.

Whether thirty-five cents is enough is not a question a pricing site can answer for you, and it depends almost entirely on whose code it is. For a personal side project it is free money and the honest recommendation is to take it. For a company, note that the sum involved is $98 per developer per month, which is less than the seat price of most of the tools those developers already use, and weigh it against a permission whose exact scope Meta has not written down. The asymmetry is not really about the money. It is that the buyer has priced the asset and the seller mostly has not.

The rate limits describe the data Meta is shopping for

The rate limits have mostly been reported as a catch, which they are. Read the two caps against each other instead and they say something more specific.

Standard gets 3,000 requests per minute and 4 million tokens per minute per team. Contributor gets 60 requests per minute and 2.1 million tokens per minute, plus a separate cap of 600 background submissions per minute. So the request allowance was cut by 50x while the token allowance fell by less than half. Those are not the same restriction wearing two hats. Divide one cap by the other and you get the request each tier is provisioned for: the standard tier budgets about 1,333 tokens per request, the contributor tier about 35,000. Twenty-six times larger.

Read that alongside the near-total cache discount and the two point the same way. A tier that wanted volume would have cut requests and tokens together. This one is configured to accept a small number of very large calls, and to make repeated large prefixes nearly free. The thing that is cheapest to send is a whole repository, over and over. Whatever Meta's intent, that is what the price and the limits jointly select for, and it is worth knowing before you point a coding agent at a private monorepo.

One consequence for anyone treating this as a production option: 60 requests per minute is one per second, shared across a team. That is fine for a handful of developers running an agent, and it rules out serving the tier to end users. The ceiling on the whole arrangement, if you saturated the token cap through an eight-hour day for twenty-two days, is around 22 billion tokens and roughly $25,500 of discount a month. Nobody is going to reach that. It is just useful to know the offer is bounded.

The price you are being sold is not a floor

Here is where the trade gets harder to justify, and it is a straightforward matter of looking at the rest of the market. If $0.10 and $0.20 were unobtainable otherwise, the permission would be buying you something real. It is not.

The two middle rows are the story, and they need one honesty caveat that we would be hypocrites to omit after the section above. Poolside's Laguna S 2.1 has no first-party rate card either. The $0.10 and $0.20 is the modal price across the resellers hosting it, and OpenRouter has since moved to $0.09 and $0.18, which is cheaper still. So the precise claim is not that Poolside publishes this rate. It is that you can buy Laguna S 2.1 today, on both legs, at the number Meta is asking for your prompts, from several vendors, without a data condition. Two further caveats: the million-token context holds on the paid BF16 endpoint, not the quantized or free ones, which cap at 262,144. And Poolside's free endpoint may itself train on your traffic, which makes it a version of the same bargain rather than an escape from it.

Above it, two models are cheaper still, with caveats that cut against them. Qwen3.7 Flash blends to 44% of the contributor rate, but that $0.03 applies only below 32K of input. Past it the whole request reprices to $0.10 and $0.40, and past 256K again to $0.20 and $0.80, which is twice Meta on input and four times it on output. For a coding agent holding a large repo in context, the cheap Qwen tier is emphatically not the tier you are on, and the top tier loses to Meta outright. GPT-OSS 120B at 61% has no such cliff but caps at 131K of context, which rules it out for the same workload. The Hunyuan row is carried with a warning: that $0.066 is a reseller listing rather than a Tencent rate card, and Tencent Cloud direct is roughly 2.5x it.

None of this makes Muse Spark 1.2 a bad model, and price is not quality. The point is narrower and it is the one the launch coverage missed: the contributor tier is being presented as access to a price you could not otherwise reach, and it is not. It is a normal 2026 price for a cheap long-context model, with a permission attached that the models beside it do not ask for. You can put your own volumes through the cost calculator or line the rate cards up on the comparison pages.

On the benchmarks, briefly, because they are a mess

Muse Spark 1.2 is reported at 82.9% on Terminal-Bench 2.1, 59.3% on DeepSWE v1.1 and 70.6% on a Meta-internal bench of 440 tasks drawn from real pull requests. All three are vendor-run. The first two ran inside Meta's own Muse Code harness; the internal bench used a separate agentic harness with dedicated grading containers, which Meta's methodology report does disclose.

We are not going to rank the field on that basis, because the trackers do not agree with each other. On Terminal-Bench 2.1 Vals AI has GPT-5.6 Sol at 85.77% and Claude Opus 5 at 84.64%, Artificial Analysis puts both near 89%, and a third puts Qwen3.8-Max top at 86.6%, though that figure is itself an Alibaba self-report. A fourth, reading the official leaderboard, says none of these models have a verified entry at all. Meta's comparison set runs Muse Spark 1.1, Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash and Kimi K3, and omits GPT-5.6 Sol, which several outlets noticed. Somewhere in there is a real ordering, and the spread between published leaderboards is currently wider than the gap between the models. Treat 82.9% as a vendor claim in the neighbourhood of the frontier, and if cost per solved task is what you are after, our coding model pricing roundup covers the field on rates we can verify.

Who should take the deal

• Side projects, learning, open source you already publish: take the contributor tier. The code is public or worthless to protect, and $17 to $98 a month is a real saving for an individual.
• Anything proprietary: do not route it through a tier whose permission scope is undocumented, when the same $0.10 and $0.20 is available for Laguna S 2.1 from several hosts, with no data condition on the paid endpoint and the same million-token context.
• Mixed repos: the two model IDs differ by a suffix, so splitting traffic is a config change rather than a migration. That is the genuinely nice part of this design and it deserves credit.
• Anything user-facing: 60 requests per minute settles it. This is not a production tier and was not built to be one.
• Before committing a team: ask Meta in writing whether the grant is training only, and whether it covers evaluation and analytics. If the answer is not in a document you can keep, that is your answer.

The broader thing worth flagging is the shape of the offer rather than its novelty. Free endpoints that train on your traffic are already common, Poolside's among them, and they have always been a data trade with the price left implicit. What is different here is that Meta put a second SKU beside the first and let you read the difference, which turns an implicit trade into a quotable number. That is more honest than the usual arrangement, where training rights are taken quietly in terms of service at no discount at all. We would rather have the number than not. It just needs reading as a bid rather than a sale, and the bid is thirty-five cents a million tokens.

How we checked this, and what we could not

• Meta: introducing Muse Code and Muse Spark 1.2 - The August 5 announcement. Source for Muse Code being a terminal coding agent with async background agents, a replay-exact local event log, and the bundled /plan, /grill and /goal skills. Contains no prices, no rate limits, no context window and no data policy, and its benchmark charts are unlabelled bars. Every dollar figure in this post had to come from elsewhere
• Meta: Muse Spark 1.2 evaluation methodology - The three-page report linked from the announcement, and the reason this post does not accuse Meta of documenting nothing. Source for the comparator list that omits GPT-5.6 Sol, for pass@1 across five attempts, for the 440-task internal bench running on a separate harness with dedicated grading containers rather than Muse Code, and for Meta's own admission that its setup may not be tuned for proprietary third-party models. Contains no pricing of any kind
• Meta: Muse Code developer product page - Checked because the pricing had to be somewhere first-party. It is not here either
• Vorp Labs: Muse Spark 1.2 release review - The most complete secondary rate card, and the source for the figures this post is built on: $1.25 / $0.15 / $4.25 standard against $0.10 / $0.002 / $0.20 contributor, 1,048,576-token context on both, web search grounding at $2.50 per 1,000 queries, 3,000 RPM and 4M TPM against 60 RPM and 2.1M TPM with a separate 600-per-minute background submission cap, and the confirmation that standard rates are identical to Muse Spark 1.1
• Implicator: Meta's 21x discount for developer data - Independent corroboration of all six price points and the 3,000-against-60 RPM split, plus the data policy framing: permission to train on prompts and completions on the contributor tier, and a commitment not to on the standard tier
• Layer3 Labs: Muse Spark 1.2 standard vs contributor pricing - A third independent match on all six prices, states the figures come from Meta's launch email, and is the source for the open question this post could not resolve: whether the contributor grant is training only, or also covers evaluation, red-teaming and product analytics. Note it also states Meta disclosed no rate limits, context window or output caps at launch, which is why this post treats those figures as resting on thinner ground than the prices
• Vercel AI Gateway: muse-spark-1.2-contributor - A reseller carrying the contributor SKU at $0.10 and $0.20 with the 1,048,576-token context, and stating plainly that inputs and outputs are used to train and improve Meta's models. Useful because it is a rate card someone is actually billing against
• BenchLM: Muse Spark 1.2 benchmarks - Source for Terminal-Bench 2.1 at 82.9% and DeepSWE at 59.3% being flagged provider-run rather than independently reproduced, and for the best verified Terminal-Bench 2.1 result going to Qwen3.8-Max at 86.6%. Its ordering disagrees with other trackers, which is why this post declines to rank the field
• Our own pricing data - Comparison rates for Poolside Laguna S 2.1, Qwen3.7 Flash, GPT-OSS 120B, Hunyuan HY3 Preview, Llama 4 Scout, GLM-4.7-FlashX and DeepSeek V4-Flash come from the TokenCost rate card, which also holds the Muse Spark 1.1 entry at $1.25 / $4.25 used to verify that standard pricing did not move. Two of those rows carry warnings in our own database and carry them here too: the Laguna S 2.1 figure is a reseller price rather than a Poolside rate card, and the Hunyuan figure is a reseller listing against a Tencent direct rate roughly 2.5x higher
• What we could not verify - Whether the prices appear on any publicly reachable Meta-owned page, though one source says it checked Meta's Model API documentation behind the console; the exact legal scope of the contributor data grant; whether the 600-per-minute background cap applies to the contributor tier alone or to both; whether web search grounding is billed on the contributor tier at all; maximum output tokens; and whether enterprise accounts are eligible. The three workload profiles are our assumptions, clearly marked as such, not measurements of Muse Code