Meta Offers Discounts to Users Sharing Data with New AI Model

Meta has launched a new pricing strategy for its Muse Spark AI model, which is designed to support coding and other agent-based tasks. The company is offering significant discounts—around 95%—to users who agree to share their input prompts and model outputs to aid in the development of future AI models. Under this contributor pricing, 1 million input tokens cost just 10 cents compared to the standard $1.25, while output tokens are priced at 20 cents per million instead of $4.25. This move follows Meta's earlier, controversial attempt to collect training data by monitoring employee computer usage, which was paused after internal backlash. The company has not publicly commented on the new pricing model. Access to user data is crucial for improving AI agents. For example, the coding agent Claude Code saw significant performance gains after it began storing user sessions for reinforcement learning. However, many professional workflows lack clear digital traces, making it difficult to gather data for training outside software engineering. Princeton computer science professor Arvind Narayanan noted that large companies often avoid sharing data for model training, preferring enterprise plans that emphasize data retention and IT governance despite higher costs. Meta's contributor pricing could encourage organizations to differentiate between proprietary data and information that can be shared, potentially increasing data availability for AI development. This pricing approach also comes amid heightened competition among AI developers. Recently, Anthropic reduced costs for processing cached tokens in its new models, and OpenAI implemented significant price cuts for its latest offerings. Meta's strategy reflects a broader industry trend toward flexible pricing tied to data sharing agreements, aiming to balance user privacy concerns with the need for high-quality training data.