Two Rows, Two Prices, Five Words
On August 5, 2026, Meta Superintelligence Labs shipped its first commercial product: Muse Code, an agent that lives in your terminal — one line to install, macOS and Linux only, a million tokens of context, running on a model called Muse Spark 1.2. MSL was founded in June 2025; Alexandr Wang runs it as Meta's first Chief AI Officer, alongside Nat Friedman on products and Daniel Gross. The standard price undercut everyone: $1.25 per million tokens in, $4.25 per million out, 15 cents for cached input. A million tokens is roughly 750,000 words — call it a stack of novels going in. That price alone was the story most outlets ran.
But the pricing table has two model names, not one: muse-spark-1.2 and muse-spark-1.2-contributor. The contributor tier costs ten cents per million tokens in and twenty cents out — roughly a twelfth of a standard price that was already the cheapest thing on the market. And the entire published terms language separating the two tiers is a column label. The standard row says "Not used to improve our products." The contributor row says "Used to improve our products." Five words. As of launch week, that is the whole published policy.
The Fifty-Times Throttle Nobody Reported
The coverage fixated on the data. Almost nobody reported the other thing you pay with: time. The contributor tier is capped at sixty requests per minute. The standard tier gets three thousand requests per minute and four million tokens per minute. That is a fifty-times throttle, welded onto the discount deliberately.
Think about who lives inside a sixty-a-minute limit and doesn't care. One person. A side project at eleven at night. A two-person startup. They will never hit sixty requests a minute, so the throttle is invisible to them — it costs them nothing. And the throttle is exactly what makes the tier useless to a company big enough to have lawyers. The rate limit isn't a limitation; it sorts the customers.
The Correction: There Is No Locked Door
An early claim — one Lucy made herself, and corrects on air — was that enterprises are excluded from the cheap tier. That isn't true. Meta published no carve-out, no eligibility bar, no prohibition of any kind. The true version is an absence: no stated retention window, no stated processing region, no data processing agreement, no server-side audit log, no admin console, no single sign-on, no scoped compliance attestation. A bank, a hospital, a defence contractor — they don't need permission to use the contributor tier. They need a document that doesn't exist. It's not a locked door; there's simply nothing to sign. A rule you can point at. An absence you can't — which is arguably worse.
Wrong by a Factor of Ten
A figure circulated in the coverage: that the cheap tier is "roughly 25 percent of what OpenAI and Anthropic charge." It's wrong by about a factor of ten. From the vendors' own pricing pages: Claude Opus 5 is $5 per million in, $25 out; OpenAI's coding flagship gpt-5.3-codex is $1.75 in, $14 out; gpt-5.6-sol is $5 in, $30 out. Against Opus 5, the contributor tier's ten cents is 2 percent of the input price, and its twenty cents is 0.8 percent of the output price. Blend usage the way real usage blends — 80 percent in, 20 percent out — and the contributor tier costs about 12 cents per million against roughly $9. That's 1.3 percent. Not a quarter of the price. A hundredth.
The mistake itself is the interesting part. Take the standard tier's $1.25 against Opus 5's $5 input and you get exactly 25 percent. Someone took a true sentence about the standard tier and set it down next to the wrong paragraph. The number was real; it just wandered. Two caveats in the honest column: Anthropic's own documentation says that from Claude 4.7 onward its tokenizer produces about 30 percent more tokens for the same text, so every per-token comparison in this story flatters Muse by an unknown margin — nowhere near enough to close a hundred-to-one gap, but real. And Claude Sonnet 5 at $2/$10 is introductory pricing through August 31, after which it's $3/$15.
The Chart Meta Loses, and the Agent That Invented Anti-Gravity
Meta's own launch materials include benchmark comparisons on Terminal-Bench 2.1, DeepSWE 1.1, and an internal test — and on all three, by Meta's own graphics, Claude Opus 5 beats Muse Spark 1.2. They drew the chart, looked at it, and shipped it. That reads as positioning, not accident: they're not claiming to be better, they're claiming to be enough, at a hundredth of the price. The small print is less flattering — they benchmarked against OpenAI's midline gpt-5.6-terra rather than the top-tier gpt-5.6-sol, and reviewers noted that part of the 6.7-point Terminal-Bench gain reflects the new agent harness rather than the model.
Then there's the live test. A developer pointed the contributor tier at a repository and asked it to audit 222 open pull requests. It finished in under five minutes, for ten cents — the same job cost $32 on Claude — at around 190 tokens per second against 30 to 50. Then he asked it to do something harder: integrate Muse Code itself as a provider in his tool. It spent three minutes researching a Google project called "anti-gravity," which does not exist and never existed, designed its entire integration plan around the fabrication, and when confronted it agreed, apologised, and could not recover. It shipped broken code. His verdict: you cannot trust it for longer-running things. Brilliant for high-volume, low-stakes work — which is, with grim precision, exactly the profile of work the cheap tier's target customer has a lot of.
Scrip, Scale AI, and Who Can't Say No
The shape has an older rhyme. In coal and lumber towns, companies paid workers in scrip — paper or brass tokens stamped with the company's name, real money with a shape to it, spendable in exactly one direction: the company store. Britain's Truck Acts, going back to 1831, made payment in kind illegal precisely because the trick worked so well; "I owe my soul to the company store" — Sixteen Tons, 1946 — is still being quoted eighty years on. A discount is a currency too. You can't bank it, can't take it elsewhere, can't hand it to your landlord, and you can only earn it by continuing to work inside the system that issues it. The comparison isn't clean — nobody is compelled, and a developer can walk to another vendor tomorrow. But if the alternative is a hundred times the price, the honest question isn't whether they're free to leave. It's whether leaving costs so much that freedom stops doing any work.
And Meta already told the world, in dollars, what this raw material is worth: $14.3 billion for a 49 percent non-voting stake in Scale AI — closed June 2025, valuing Scale at $29 billion — a company whose business was organising and labelling training data. The contributor tier is the same purchase, restructured: instead of buying a company that gathers data, pay the people who make it, in discount. It's not a price cut; it's a procurement strategy. And stop and look at what's being bought. Not oil, not ore, not land — the raw material here is what people type while they're thinking. The half-formed attempt. The thing tried at eleven at night that didn't work. That small annoyed act of thinking is now worth enough that a trillion-dollar company built a second pricing tier to be near it.
So who can't say no? Start with who can. A large company with a compliance department looks at the tier, finds no DPA, no retention window, no audit log — and stops. Not out of virtue; out of paperwork. Their bureaucracy is, for once, doing something useful. The people not protected are one person with an idea and a laptop, a two-person startup burning savings, a student, someone in a country where $32 is not a rounding error. For them the choice isn't "cheap versus principled" — it's "build the thing, or don't." The episode is careful not to call this villainy: two rows, two prices, one sentence of difference is more legible than most of what's been done with data in the last twenty years. For years the answer to "they're harvesting your work" was "calm down, that's paranoid" — and this week Meta put it in a table, with a price on it. The paranoid position turned out to be the conservative one. What actually rankles is that the trade is offered hardest to the people least able to refuse it, and then described as a choice.
Loose ends, marked as such: around fifteen developers in Canada, the UK, Australia, and Thailand report the contributor model returning "Model not found," US VPNs don't help, and Meta has said nothing beyond "expanded global access" — reported, not settled. The Hacker News launch thread (278 points, 177 comments) complained mostly about the missing hard spending cap, Docker sign-in failures, selfie verification, and the absent Windows build. And asked whether Muse Spark would be open-sourced, Zuckerberg said only: "I'll have more to share on that soon." That's not a yes — and it's a long way from Llama.