Readers who let their prompts and outputs feed future model training can now cut their bill on a frontier coding and agent model by roughly 95%. Meta’s new Muse Spark, a model aimed at running coding and other agents, prices a million input tokens at ten cents on a contributor tier versus $1.25 on the standard rate, and output tokens at twenty cents per million instead of $4.25. The same arrangement doubles as a workaround for a long-running shortage of high-quality training data, especially for the messy workflows that fall outside software engineering.<\/p>\n
The contributor tier is essentially a usage discount in exchange for letting Meta see what users do. Under the standard agreement, one million input tokens cost $1.25 and one million output tokens cost $4.25. On the contributor tier those figures drop to $0.10 and $0.20 per million, a cut of around 92% for input and 95% for output. Meta’s published pricing guide frames the discount as a way to lower the barrier to prototyping, integration testing, and scaling experiments in which training on the resulting data is acceptable.<\/p>\n
For an individual developer running agentic coding tasks, that math turns token-heavy workloads into something close to a free trial. For a company, it raises a sharper question: which of its prompts, outputs, and internal traces are genuinely proprietary, and which could safely be handed over for a discount that turns a six-figure inference bill into a five-figure one.<\/p>\n
Agentic tools, models that can plan multi-step tasks, call other software, and write or edit code on a user’s behalf, only improve when developers can watch them fail and succeed on real work. The kind of traces left by a chat with an assistant is thin. The traces left by a coding agent, every file opened, every command run, every test passed or broken, are far richer, which is why coding has become the proving ground for agent capability.<\/p>\n
Meta has had a rough time collecting that kind of signal at home. An internal initiative launched earlier this year to track the computer usage of its own employees attracted wide internal criticism and was paused in June. The company did not respond to questions about the new external pricing scheme.<\/p>\n
The technical reason the data matters is that reinforcement learning on real agent trajectories is one of the few methods that reliably improves long-horizon tool use. A widely cited example of the payoff is the jump in coding agent capability between April 2025 and October 2025, attributed to one tool storing all coding agent sessions by default and reusing them as reinforcement learning training data. That single decision gave the underlying model a much larger and much messier set of successful and failed multi-step runs than any hand-curated benchmark could supply.<\/p>\n
Enterprises that already pay frontier labs tend to stay on token-billed enterprise plans even when subscription-based consumer plans are discounted by ten to twenty times or more. The premium is not buying better answers. It buys control over retention and IT governance, and it keeps proprietary prompts and outputs out of training pipelines.<\/p>\n
That dynamic is exactly what the Muse Spark contributor tier tries to crack. Instead of asking firms to donate data altruistically, Meta puts a price on it. If a discount of roughly 95% applies to a workload whose prompts and outputs are not trade secrets, the rational move is to opt in. The model gets more trajectories, the user gets a near-zero inference bill, and the data flows both ways.<\/p>\n
One side effect is that procurement, legal, and security teams will have to draw sharper lines around what counts as proprietary. Anything not flagged as sensitive becomes a candidate for the contributor tier, which effectively converts an opaque training-data decision into a per-token cost decision.<\/p>\n
The Muse Spark structure lands in the middle of an aggressive round of price cuts across the frontier labs. Anthropic’s newest Fable and Mythos models arrived with reduced prices on cached tokens, the repeat prompts an agent sends within a single conversation, where reuse is heaviest. OpenAI’s latest models got major price cuts at the end of July. Meta’s move is structurally different: rather than cutting list prices, Meta is cutting the price only for users who accept training on their data.<\/p>\n
That framing gives model builders a way to compete on price without losing access to the training signal they need for agent improvements. It also lets them charge full price to enterprise customers who need strict retention controls, and a deeply discounted price to users for whom the data is the point.<\/p>\n
For anyone building or buying agentic tools, the practical takeaway is that the cost of inference is decoupling from the cost of data. The cheapest way to run an agent will increasingly depend on what the provider gets back in return, and on how clearly an organization can separate its sensitive workflows from the ones it can afford to share.<\/p>\n
Muse Spark is Meta’s new model aimed at operating coding and other agents. It is sold under two pricing tiers, a standard rate and a contributor rate that discounts usage in exchange for sharing prompts and outputs for future model training.<\/p>\n
The contributor tier prices one million input tokens at $0.10 and one million output tokens at $0.20, compared with $1.25 and $4.25 respectively under the standard agreement. That works out to an average discount of roughly 95%.<\/p>\n
Agentic tools leave richer traces than chat assistants do, including files opened, commands run, and tests passed or broken, which makes their trajectories especially useful for reinforcement learning. Improvements to coding agent capability between April 2025 and October 2025 are widely attributed to one tool storing coding agent sessions by default and reusing them for training.<\/p>\n