Contributor pricing is Meta’s name for a simple trade: hand over your prompts and the model’s replies for training, and your bill drops by up to 21 times. TechCrunch’s Tim Fernholz reported the tier on 3 September 2026, a day after Meta shipped Muse Spark 1.3, and framed it in one line. Most AI tools let you opt out of sharing your usage. Meta has taken that idea and put a price tag on it.
The numbers are not subtle. A million input tokens costs $1.25 on the standard agreement and 10 cents under contributor pricing. A million output tokens costs $4.25 standard and 20 cents contributor. Meta’s own pricing guide says the tier “lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable.” The company declined to comment to TechCrunch on the new model.
This article walks through what the discount actually buys, the arithmetic on a real agent workload, why Meta wants this data badly enough to pay for it, and the throughput penalty nobody mentions. It ends with the classification and routing work any business needs before letting a single request touch the cheap tier, including the AI agents most teams are now running unsupervised.
Table of contents
- What Meta’s Contributor Pricing Actually Buys
- The Arithmetic That Makes Contributor Pricing Look Irresistible
- Why Contributor Pricing Exists: Meta Needs Your Prompts
- Contributor Pricing Against the Rest of the Market
- What Muse Spark 1.3 Delivers Under Contributor Pricing
- The Hidden Cost in Contributor Pricing: Throughput
- Who Should Never Put Traffic on Contributor Pricing
- Who Can Take Contributor Pricing Safely
- The Governance Work Contributor Pricing Forces On You
- Contributor Pricing and UK Data Protection Law
- What Contributor Pricing Signals About the Next Two Years
- Contributor Pricing FAQ
- References
What Meta's Contributor Pricing Actually Buys
The first thing to understand about contributor pricing is what it is not. It is not a smaller model, a distilled variant, or a quantised copy running on cheaper hardware.
Two SKUs, one checkpoint
Meta exposes two model identifiers, muse-spark-1.3 and muse-spark-1.3-contributor. They point at the same weights, the same checkpoint and the same 1,048,576-token context window, with a maximum completion length of 943,718 tokens. Nothing about the model changes when you switch. The only thing that changes is what Meta is allowed to do with the traffic afterwards, and what you pay for it. Contributor pricing is a commercial distinction dressed as a product one.
The exact figures
Standard access runs $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached input tokens. Contributor pricing charges $0.10, $0.20 and $0.002 for the same three lines. Cached input is the most extreme of the three at 75 times cheaper, which matters enormously for agent loops that replay a long system prompt on every turn.
What “contribute” means in Meta’s own words
Meta’s documentation describes the tier as heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models. Send traffic through the contributor SKU and Meta may train on it. Send it through the standard SKU and it will not. What the documentation does not spell out is retention periods, human review, deletion rights, or whether file attachments and tool arguments count as trainable material.
Here is the price gap in one view, per million tokens, with the standard output price as the top of the scale.
The full side-by-side
The two tiers differ on more lines than the price sheet suggests, and the throughput rows are the ones that catch teams out later.
| Line | Standard SKU | Contributor SKU | Gap |
|---|---|---|---|
| Input per million tokens | $1.25 | $0.10 | 12.5x cheaper |
| Output per million tokens | $4.25 | $0.20 | 21.25x cheaper |
| Cached input per million | $0.15 | $0.002 | 75x cheaper |
| Requests per minute | 3,000 | 100 | 30x lower |
| Tokens per minute | 4,000,000 | 3,000,000 | 25% lower |
| Model weights | muse-spark-1.3 | Identical checkpoint | No difference |
| Context window | 1,048,576 tokens | 1,048,576 tokens | No difference |
| Training rights over your traffic | Meta will not train | Meta may train | The whole trade |
The Arithmetic That Makes Contributor Pricing Look Irresistible
Percentages hide how large this gap becomes at production volume. Put a plausible workload through both tiers and the difference stops being a line item.
A worked example on 500 million tokens
Take a mid-sized engineering group running coding agents: 500 million input tokens and 100 million output tokens in a month. On standard pricing that is 500 lots of $1.25, which is $625, plus 100 lots of $4.25, which is $425. The monthly bill is $1,050. Under contributor pricing the same traffic costs 500 lots of $0.10, which is $50, plus 100 lots of $0.20, which is $20. The bill is $70.
The saving in plain terms
That is $980 saved on $1,050, or 93.3% of the bill. TechCrunch put the average discount at about 95%, and a workload weighted more heavily toward output lands closer to that figure, because output carries the 21.25x multiple rather than the 12.5x one. The more your agents talk, the better contributor pricing looks.
The split-routing middle path
Almost nobody should run everything on one tier. If 70% of that traffic is safe to share and 30% is not, the bill becomes 70% of $70 plus 30% of $1,050, which is $49 plus $315, or $364. Here are the three options side by side, with the standard bill as the top of the scale.
Why the mixed number is the honest one
The mixed figure is the one to plan against, because it prices the reality that some of your traffic carries client names, staff records or unreleased product detail. A team that budgets on the $70 line and then discovers half its workload cannot legally be shared has not saved money. It has built a migration it will have to unwind.
Why Contributor Pricing Exists: Meta Needs Your Prompts
Meta is not running a charity drive on inference costs. Contributor pricing is a data acquisition programme with a billing system attached, and Meta arrived at it after a more direct approach failed publicly.
The Model Capability Initiative and how it collapsed
In April 2026 Meta announced the Model Capability Initiative, an internal programme that captured mouse movements, keystrokes and occasional screenshots from company-issued devices in the United States, to teach models how humans actually navigate a computer. Employees could not opt out. More than 1,500 of them signed a petition calling it an Employee Data Extraction Factory.
What actually stopped it
The petition did not stop it. An internal leak did. Sensitive material gathered by the tool, reportedly including private conversations, performance information and meeting transcriptions, became visible across the company. Meta classified the incident as a SEV 2 on its internal severity scale and paused the programme on 23 June 2026. TechCrunch noted the pause in its contributor pricing story as evidence that Meta has had a rough time obtaining training data.
Coding agents got good because sessions were kept
The value of this traffic is not theoretical. Mario Zechner, the developer behind the open-source harness Pi, told TechCrunch that the jump in coding-agent capability between April 2025 and October 2025 came from Claude Code storing sessions by default and using them for reinforcement learning training. Real transcripts of real agents doing real work turned out to be the scarce ingredient, not compute.
Everything outside code has no trace
Zechner also explained why progress stalls outside software. Everything is coding-agent shaped, he argued, because that is the only place the training data exists. His example is a manager who makes a decision today whose outcome lands months later. That cannot be captured in a simple back-and-forth trace, so anything a business has not digitised is invisible to a model trying to learn it. Contributor pricing is an attempt to buy those missing traces from the outside.
The enterprise objection Meta is trying to price around
Arvind Narayanan, a computer science professor at Princeton, has pointed out that large companies already reveal their preference here. They stick with token-billed enterprise plans even though consumer subscriptions like Claude Max and ChatGPT Pro are discounted by 10 to 20 times or more, and the main difference between the plans is data retention plus enterprise IT governance. Firms are already paying a large premium not to be trained on. Contributor pricing simply makes the size of that premium explicit, and Narayanan suggested it could push companies to work out which of their data is genuinely proprietary and which is not.
Contributor Pricing Against the Rest of the Market
The tier landed in the middle of a price war, which is part of why it reads as aggressive rather than desperate.
The week it launched into
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 2 September 2026 with lowered costs for processing cached tokens. OpenAI cut prices across its latest models at the end of July 2026. Meta shipped Muse Spark 1.3 on 2 September and the contributor tier surfaced the next day. Mark Zuckerberg had already set the tone in July, saying the pricing from some of the other labs is very extreme and has very high margins, and promising that Meta’s would be very aggressive and attractive.
List prices side by side
Meta’s standard tier already undercuts the frontier competition before the contributor discount is applied. Against published list prices, the contributor tier is in a different category entirely.
| Model | Input per 1M | Output per 1M | Trains on your API traffic |
|---|---|---|---|
| Muse Spark 1.3 contributor | $0.10 | $0.20 | Yes, by design |
| Muse Spark 1.3 standard | $1.25 | $4.25 | No |
| Claude Haiku 4.5 | $1.00 | $5.00 | None by default |
| Claude Sonnet 5 | $2.00 | $10.00 | None by default |
| Claude Opus 5 | $5.00 | $25.00 | None by default |
| Claude Fable 5.1 | $10.00 | $50.00 | None by default |
Reading the table honestly
List price is not the whole cost of an API. Cache behaviour, batch discounts, retry rates and how many tokens a model burns to reach an answer all move the real figure. But the column that matters for this article is the last one. Every other row treats not training on your traffic as the default. Contributor pricing is the first mainstream tier to treat it as a chargeable upgrade.
What Muse Spark 1.3 Delivers Under Contributor Pricing
A cheap model nobody wants to use is not a bargain. The reason contributor pricing is interesting is that the underlying model is genuinely competitive.
The published benchmarks
Meta reports the shipping configuration at 1,709 Elo on GDPval-AA v2, 57.2 on OSWorld 2.0, 61.2 on JobBench, 89.2 on Terminal-Bench 2.1 and 89.4 on DeepSearchQA. Artificial Analysis scores it at 61 on its Intelligence Index. Zuckerberg described the release as frontier performance almost too cheap to meter, and Alexandr Wang posted a pointed line about Gemini.
Where that sits against rivals
On the same Artificial Analysis index, Claude Opus 5 scores 63, GPT-5.6 Sol scores 62, Muse Spark 1.3 scores 61 and Gemini 3.8 Flash scores 59. Gemini is faster in raw throughput at 305 tokens per second against 235. The gap at the top is four index points across the whole field.
The tier you cannot buy yet
Meta’s strongest published numbers come from a maximum reasoning configuration that is still finishing safety testing, reported at 1,754 Elo on GDPval-AA v2, 66.9 on OSWorld 2.0 and 64.9 on JobBench. That configuration is not broadly available, so the benchmarks most likely to be quoted at you in a sales conversation are not the ones you can call today. Judge contributor pricing on the shipping model, not the pending one.
The price sheet is only half the comparison, and the half everyone reads. The rate limits are where the cheap tier stops being free money.
Thirty times fewer requests per minute
Standard access allows 3,000 requests per minute. Contributor pricing allows 100. Tokens per minute drop more gently, from 4,000,000 to 3,000,000. For a batch job that sends a few enormous requests, the token ceiling is what binds and the tier is close to a straight discount. For an agent that fires hundreds of small tool calls a minute, 100 requests per minute is a wall.
What that does to agent architectures
Modern agent loops are chatty by design. A single task might involve a plan, a dozen tool calls, several retries and a summarisation pass. Under contributor pricing, a handful of concurrent users can saturate the request ceiling while barely touching the token ceiling. The workloads that benefit most from the price cut are often the ones least able to live inside the limit.
The realistic pattern
Most teams that adopt this will end up holding two keys and routing between them, which is exactly what the mixed-bill arithmetic assumed. Treat contributor pricing as a lane for high-volume, low-concurrency, non-sensitive work, and keep standard access for interactive traffic that has to answer now.
Who Should Never Put Traffic on Contributor Pricing
Some categories are straightforward no. Not because the tier is badly built, but because the permission you are granting is broad and the material is not yours to grant it over.
Anything covered by a client confidentiality clause
If you are a professional services firm, an agency, an accountancy practice or a managed service provider, your client contracts almost certainly restrict onward disclosure. Routing client material through a tier whose entire purpose is training somebody else’s model is a disclosure. The discount will not cover the conversation with the client afterwards.
Personal data of any kind
Names, contact details, HR records, customer support transcripts and anything else identifying a living person belong on the standard tier. Consent and lawful basis are not solved by a cheaper invoice, and data protection obligations do not pause because a vendor offered a discount.
Unreleased product and commercial strategy
Pricing models, roadmaps, merger material, source code for unpublished products and anything under embargo are exactly the traces a competitor would value most. The risk is not that a future model recites your roadmap verbatim. It is that you have handed a broad training permission over your most differentiating material for a saving measured in hundreds of dollars.
Regulated and safety-critical content
Health, financial advice, legal advice and anything sitting inside a regulated process carry their own retention and audit rules. Where the documentation is silent on retention periods, human review and deletion rights, silence is not an answer you can put in front of a regulator.
Who Can Take Contributor Pricing Safely
The tier is genuinely useful, and dismissing it wholesale is as lazy as adopting it wholesale. Whole categories of work carry no confidentiality weight at all.
Prototypes, spikes and evaluations
Building a proof of concept against synthetic fixtures, benchmarking three prompt strategies, or running an evaluation harness over public datasets involves nothing you would mind a model learning. This is exactly the use Meta’s guide names, and contributor pricing turns an experiment budget into a rounding error.
Public and published material
Summarising your own published blog posts, parsing open government data, processing public filings or working over open-source code you already released puts nothing new into the world. The material is already in the training set of every frontier model.
Synthetic and load-test traffic
Load tests, regression suites and generated fixtures cost real money at production volume and contain nothing. Moving them onto contributor pricing is close to a free saving, and the 100-requests-per-minute ceiling matters less when the schedule is yours to set.
Internal tooling over non-sensitive corpora
Documentation search over public docs, internal tooling that touches no customer records, and developer utilities that operate on open dependencies are all reasonable candidates. The test is not whether the workload is internal. It is whether the content would embarrass you if it surfaced in a model’s behaviour two years from now.
| Workload | Tier | Reason |
|---|---|---|
| Client deliverables and correspondence | Standard only | Confidentiality clauses forbid onward disclosure |
| Anything with personal data | Standard only | Lawful basis and retention rules apply |
| Unreleased roadmap or source code | Standard only | Your most differentiating material |
| Regulated or safety-critical content | Standard only | Retention and audit obligations |
| Prototypes on synthetic fixtures | Contributor | No real content involved |
| Already-published material | Contributor | Nothing new is disclosed |
| Load tests and regression suites | Contributor | Generated content, schedulable around limits |
| Interactive user-facing chat | Standard | Request ceiling and unpredictable input |
The Governance Work Contributor Pricing Forces On You
The uncomfortable part of this tier is that using it properly requires capability most teams do not have yet. You cannot route by sensitivity if you have never classified anything.
Classify before you route
Every workload needs a label before it can be assigned a tier, and the label has to live somewhere a router can read it. That is ordinary data management work, and it is the step teams skip. A rule that exists only in a policy document cannot stop a request at runtime.
Two keys, two SKUs, one router
The mechanical control is simple. Issue separate credentials for each tier, never let application code choose between them, and put the decision in a gateway that reads the workload’s classification. Developers should not be able to move traffic to the cheap tier by editing a model string, because at some point one of them will, on a deadline, without thinking.
Log which tier served which request
Retrospective questions are the ones that hurt. When a client asks whether their material was ever used for training, the only acceptable answer is a query against a log, not a recollection. Record the SKU alongside every request and keep those records for as long as your client contracts run. This is the same discipline good IT governance applies to every other data flow.
Review the boundary on a schedule
Classification decays. A prototype that used synthetic fixtures in March is running against real customer records by September, and nobody moved it back to the standard tier. Put contributor pricing routing on the same review cycle as access rights, and check what is actually flowing rather than what was approved.
| Control | What it prevents | Effort |
|---|---|---|
| Workload classification register | Routing decisions made by guesswork | Days, once |
| Separate credentials per tier | Accidental promotion of sensitive traffic | Hours |
| Gateway-side tier selection | Model strings edited under deadline | Days |
| Per-request SKU logging | Unanswerable client questions later | Hours |
| Scheduled boundary review | Classification drift over months | Quarterly |
| Contract and DPA check | Breaching a client confidentiality clause | Days, once |
Contributor Pricing and UK Data Protection Law
For a UK business the analysis does not stop at commercial preference. Deliberately routing traffic into a training pipeline is a processing decision with legal weight.
You are choosing the purpose, so you are the controller
If your organisation decides that a set of prompts will be used to train a third party’s model, that is your decision about purpose. Under UK GDPR you need a lawful basis for it, and the basis has to exist before the traffic flows, not after somebody asks. The ICO’s guidance on lawful basis is the right starting point, and it is a short read compared with the cost of getting this wrong.
The processor relationship changes shape
Standard API access looks like a straightforward processor arrangement: the vendor handles data on your instructions and does not repurpose it. Contributor pricing explicitly repurposes it. That difference belongs in your record of processing activities and in any data protection impact assessment covering the system, not in a footnote on an invoice.
Contracts you signed before this tier existed
Most client contracts and data processing agreements were drafted when training on your traffic was not an option anyone was buying. Read the onward-disclosure and sub-processor clauses before enabling contributor pricing on anything client-adjacent, and expect at least some of them to require a conversation rather than a unilateral decision.
Transparency to the people involved
Where any personal data is in scope, the people it concerns have a right to know what happens to it. A privacy notice that describes AI processing in general terms will not obviously cover a deliberate choice to contribute their material to a model’s training set. Compliance here is cheaper to design in than to retrofit, which is the same lesson every AI strategy engagement teaches eventually.
What Contributor Pricing Signals About the Next Two Years
Whatever you decide about this specific tier, the pattern behind it is going to spread, because the economics work for both sides.
Data is being unbundled from compute
Until now the price of an API call bundled the model, the hardware and an implicit promise about your data. Contributor pricing separates that last part and puts a number on it. Narayanan’s point cuts both ways: once companies can see the premium they pay for privacy, some will decide a great deal of their traffic never needed the protection.
Expect competitors to answer
A 21x gap is hard to ignore in a market where buyers compare price sheets weekly. Anthropic and OpenAI both moved on price within the last six weeks. If contributor pricing pulls meaningful volume, the shape of the response will tell you a lot about how badly every lab needs non-coding traces.
The scarce resource is your workflow, not your text
The deeper message in Zechner’s argument is that model builders have run out of easy traces. Text on the open web is exhausted. What they lack is the record of professional work being done, step by step, in tools that were never designed to emit that record. Contributor pricing is one way to buy it. Employee monitoring was another, and it did not go well.
The right posture is deliberate, not reflexive
Neither reflex serves you here. Refusing on principle leaves real savings on the table for genuinely harmless workloads. Accepting on price hands over material you cannot get back, under permissions you have not read, to fix a bill that was rarely the binding constraint. Classify first, route second, and treat contributor pricing as one lane among several rather than a company-wide switch.
Contributor Pricing FAQ
Is the contributor model weaker than the standard one?
No. Both identifiers resolve to the same checkpoint with the same 1,048,576-token context window. The difference is price, rate limits and Meta’s rights over your traffic.
How much cheaper is contributor pricing exactly?
Input is 12.5 times cheaper, output is 21.25 times cheaper and cached input is 75 times cheaper. TechCrunch summarised the blend as roughly 95%, and the exact figure for your workload depends on your input-to-output ratio.
Can I move between tiers?
They are separate model identifiers, so switching is a configuration change rather than a migration. That is precisely why the decision belongs in a gateway rather than in application code.
Does contributor pricing affect data already sent on the standard tier?
Nothing published suggests it applies retroactively. Standard traffic is described as not used for training. Keep the logs that would let you demonstrate which tier handled which request.
What does Meta not say about the tier?
The documentation does not spell out retention periods, human review practices, deletion rights, or whether file attachments and tool arguments count as trainable data. Those gaps are the reason to keep regulated and client material on standard access.
Is this legal for a UK business?
It can be, for the right workloads. It requires a lawful basis where personal data is involved, a check against client contracts, and an update to your processing records. It is not a purely commercial decision.
What should we do first?
Classify your workloads before you touch the pricing page. Without a classification register, any routing rule is guesswork, and the saving is not worth an unanswerable question from a client.
References
Meta is paying to peek at how you use their latest AI model
Meta Developer Docs: Pricing and Rate Limits
Meta says Muse Spark 1.3 has frontier performance
Meta offers discounted access to Muse Spark 1.3 model for data sharing
Muse Spark 1.3: Pricing, Specs, and the Contributor Tier
Meta pauses employee tracking AI programme after leak
Meta’s AI training with keystrokes: Progress or privacy issue
OpenAI is building an AI agent for everything. Will everyone use them?
Arvind Narayanan on enterprise plans and data retention
Advancing the price-performance frontier with GPT-5.6
Your AI Margin is Meta’s Opportunity
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.