Two prices, one model
Meta is selling access to Muse Spark, its model for coding and agent work, at two prices that differ by whether the customer lets Meta train on what they send it. TechCrunch reported the arrangement on 3 September.
On the standard tier, Muse Spark costs $1.25 per million input tokens and $4.25 per million output tokens, and the customer’s prompts and completions stay out of Meta’s training pipeline. On the Contributor tier, the same model costs $0.10 and $0.20. Output tokens are 21 times cheaper; input tokens 12 times cheaper. TechCrunch put the blended discount at roughly 95%.
What Meta takes in return is the prompts, the model’s responses and the usage patterns around them, all of which go into training future models. Meta’s own pricing documentation describes the tier as lowering “the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable”. Meta declined to comment to TechCrunch.
Why this is different from the usual arrangement
Every major lab has a position on training from customer data, and most of them express it as a default with a toggle: consumer products train by default and let you opt out, enterprise contracts do not train and cost more. The trade is real but it is diffuse, spread across a terms-of-service page and a settings screen.
Meta has done something narrower and more legible. It has put a number on the data. A developer choosing between $4.25 and $0.20 per million output tokens is being shown, in the only unit that matters to a procurement conversation, what their coding traffic is worth to a frontier lab.
The tier is also not an experiment. According to TechCrunch, Meta first tested the structure with Muse Spark 1.2 in August 2026 and carried it forward into 1.3, which the company released on 2 September.
What developers actually send
The reason coding traffic is worth this much is that it is unusually good training data. An agent session contains a task, a series of attempts, tool calls, error messages and — crucially — a signal about whether the thing worked. That is expensive to synthesise and hard to buy.
It is also, in a working developer’s hands, other people’s material. Prompts sent from a coding agent routinely contain proprietary source, internal API shapes, infrastructure details and occasionally credentials. The person clicking through to the cheaper tier is frequently not the person who owns what is in the context window.
Arvind Narayanan, a Princeton computer science professor, told TechCrunch that large enterprises tend to avoid sharing proprietary data and stay on enterprise plans despite the higher cost. That is the expected split: the discount lands hardest on small teams and individual developers, who have the least legal capacity to reason about what they are giving away and the most price sensitivity pushing them to give it.
What to watch
The interesting question is whether other labs follow. Explicit data-for-price tiers have existed at the edges — free tiers that train, paid tiers that do not — but not as a headline number on a frontier model’s API page.
If OpenAI, Anthropic or Google publish a comparable tier, the industry will have established that training data has a public unit price, and every enterprise negotiation about data handling will start from it. If none of them do, Meta’s tier will be read as what it may also be: a company that is behind on data buying its way forward.
