Meta is offering a steep discount on access to its Muse Spark model when customers agree to share their prompts and model outputs for future development. The arrangement makes the value of real user interactions explicit: customers can pay the standard price for greater control, or contribute data in exchange for substantially cheaper usage.
TechCrunch reports that the discount averages about 95 percent. It is aimed at Muse Spark users building coding systems and other agents, where traces of complete tasks can help a model provider evaluate behavior that is difficult to reproduce in a laboratory.
Lower prices come with a data choice
Under Meta’s standard agreement, one million input tokens cost $1.25 and one million output tokens cost $4.25. Customers in the contributor tier pay 10 cents for the same quantity of input and 20 cents for output. Those terms turn permission to use prompts and responses into a form of compensation delivered through lower computing charges.
Meta’s pricing guide says the contributor option is intended to reduce the cost of prototypes, integration tests and larger experiments when customers are comfortable allowing training on their data. The offer is therefore not presented as a universal default. It separates workloads where sharing is acceptable from those where information control and enterprise governance matter more.
The source does not detail how contributed material will be selected, retained or incorporated into later models. Meta also did not answer TechCrunch’s question about the pricing plan. The disclosed prices show the economic choice, but they do not by themselves resolve how customers should assess the sensitivity of individual prompts and outputs.
Agent workflows create valuable traces
The incentive reflects a practical problem for developers of agentic systems. These tools operate through sequences of actions, and observing real sessions can expose where a model succeeds, hesitates or fails. Mario Zechner, developer of the open-source Pi harness, connected an earlier improvement in coding agents to the storage of user sessions for reinforcement-learning training.
Collecting comparable evidence outside software work is harder. TechCrunch notes that many professional processes are complex and leave fewer useful digital traces. As model companies try to extend agents beyond coding, examples of real work can become especially valuable for evaluation and improvement.
Meta has already faced resistance over another data-gathering effort. An internal program designed to track employees’ computer use drew criticism and was paused in June. That episode does not establish how customers will react to Muse Spark’s voluntary tier, but it provides context for Meta’s decision to attach a visible financial benefit to participation.
Enterprise control meets price competition
Princeton computer science professor Arvind Narayanan argued that large organizations have demonstrated a willingness to pay more for plans that include data-retention controls and enterprise governance. He pointed to companies choosing token-billed enterprise products even when consumer subscriptions can be much cheaper.
Contributor pricing may force buyers to classify their information more deliberately. Narayanan suggested that explicit compensation could encourage companies to distinguish genuinely proprietary material from data they are willing to share. That is a possible consequence, not proof that organizations will adopt the tier or improve their internal controls.
The offer also arrives as frontier model providers compete on price. TechCrunch cites lower cached-token costs for Anthropic’s newest models and major OpenAI price reductions at the end of July. Meta’s approach adds a second variable to that contest: the cheapest rate depends not only on compute efficiency, but on what the customer permits the provider to learn from its work.
For buyers, the headline discount is easy to quantify while the value of surrendered workflow data is not. The plan’s significance lies in making that asymmetry visible. Meta is no longer treating user interactions merely as a policy setting; it is assigning them a price and asking enterprises to decide which side of the exchange they prefer.
