Meta Offers 95% Discount on Muse Spark in Exchange for Your Training Data

Meta has introduced a radical new pricing tier for its Muse Spark AI model: users who agree to share their prompts and model outputs with Meta for future model training receive approximately 95% off standard API costs. Input tokens drop from $1.25 to $0.10 per million, and output tokens from $4.25 to $0.20. The move comes after Meta's internal employee data-tracking initiative was paused in June following widespread criticism, and as frontier AI labs compete aggressively on pricing to lock in developer ecosystems.

Key Takeaways

Meta has introduced a radical new pricing tier for its Muse Spark AI model: users who agree to share their prompts and model outputs with Meta for future model training receive approximately 95% off standard API costs.
Input tokens drop from $1.25 to $0.10 per million, and output tokens from $4.25 to $0.20.
The move comes after Meta's internal employee data-tracking initiative was paused in June following widespread criticism, and as frontier AI labs compete aggressively on pricing to lock in developer ecosystems.

AI synthesis3 sources5

Meta has rolled out an unconventional pricing strategy for its new Muse Spark model, offering users a 95% discount on API costs in exchange for permission to use their data to train future AI models. Under the standard agreement, one million input tokens cost $1.25; under the contributor pricing model, the same million costs just $0.10. Output tokens see a similar cut: from $4.25 to $0.20 per million.

The contributor tier, as Meta describes it on its official pricing page, lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable. It is the company's most explicit attempt yet to put a price tag on the kind of user interaction data that has become critical for building capable AI agents.

The move follows a turbulent year for Meta's data acquisition efforts. An initiative launched earlier in 2026 to track the computer usage of its own employees attracted widespread internal criticism and was paused in June. The company did not respond to questions about its new pricing model.

The business logic is straightforward: data from real-world coding agent sessions has proven essential for improving AI tools. Mario Zechner, the developer behind the open-source harness Pi, told TechCrunch that the reason for a big jump in coding agent capabilities between April and October 2025 was that Claude Code would store all coding agent sessions and use them for reinforcement learning training.

But the contributor model also highlights a growing tension in the AI industry. Princeton computer science professor Arvind Narayanan noted that large companies actively resist sharing their data with model providers, sticking with token-billed enterprise plans even when consumer plans are 10-20x cheaper. The main differentiator: data retention policies. Meta's approach — paying developers to opt in — could shift that dynamic by making the trade-off explicit.

The pricing war is heating up across the frontier: Anthropic recently released its Fable and Mythos models with discounted cached token pricing, while OpenAI slashed prices on its latest models at the end of July. Meta's contributor model adds a new dimension: not just competing on capability or price, but on willingness to share the data that powers the next generation of AI.