Meta Prices Muse Spark Tiers, Offers 95% Discount for Data Contributors

Meta has introduced a differentiated pricing mechanism for its Muse Spark model series, offering customers who agree to share prompts and model responses for future training an average discount of about 95% on model calls. The latest Muse Spark 1.3 model was released on September 2, and Meta CEO Mark Zuckerberg said the model represents the company's largest improvement in coding and agent capabilities to date.

Meta has set two tiers for the Muse Spark series: "Standard" and "Contributor." Both tiers use the same model with no difference in capabilities; the key distinction is whether Meta may use user prompts and model outputs to improve its products. According to Meta, data under the standard tier will not be used for training. Meta's official pricing documentation shows that the standard tier charges $1.25 per million input tokens, while the contributor tier charges only $0.10. For output tokens, the standard price is $4.25 per million, compared with $0.20 for the contributor tier. The most significant price gap is for cached input tokens, which are portions of context that are repeatedly sent by agents and hit the cache without requiring recomputation. The standard tier charges $0.15 per million tokens, while the contributor tier charges $0.002, a difference of about 75 times.

Meta has previously faced difficulties in obtaining training data. A program launched earlier in 2026 to track employee computer usage drew widespread internal criticism and was suspended in June. User data is critical to improving the performance of agent tools, but vendors now face two obstacles in acquiring data. First, the data does not exist: as model usage shifts from software engineering to broader industry scenarios, many professional workflows are complex and leave no digital trace, making it difficult for vendors to evaluate agent performance or make improvements.

The Muse Spark pricing guide positions the contributor tier for scenarios where "you are okay with your data being used for training," such as building prototypes, testing integrations, or running batch experiments—work that is not yet in production and where data is not sensitive.