Meta has introduced a lower-cost usage plan for its newly launched Muse Spark model. Users who agree to share their prompts and model outputs for future model training can reduce their average invocation costs by approximately 95% compared to the standard plan. This turns the question of “whether to retain user data” from a privacy option into a clear price trade-off.
Shared data can be exchanged for lower prices
According to Meta’s published pricing, under the standard plan, the cost is $1.25 per 1 million input tokens; under the “contributor pricing,” it drops to $0.10. The price difference for output tokens is even greater: $4.25 per 1 million under the standard plan, and $0.20 under the contributor plan.
This means that customers willing to allow Meta to use their interaction data can test coding agents and other agent-based applications at significantly lower costs. Meta states in its pricing details that this tier lowers the barrier to prototyping, integration testing, and scaling experiments, provided customers agree to have their data used for training.
Actual usage records are more valuable.
For large model companies, real usage records are becoming increasingly important, especially for coding agents and broader agent tools. According to TechCrunch, citing Mario Zechner, developer of the open-source tool Pi, a significant improvement in coding agent capabilities by mid-2025 is linked to Claude Code’s default behavior of saving sessions and using them for reinforcement learning training.
As model vendors shift their focus away from software engineering toward specialized workflows, obtaining such data has become more difficult. Many enterprise processes are more complex and leave fewer traceable digital footprints, limiting the ability to evaluate and iterate on models.
Businesses place greater emphasis on data retention.
Princeton University computer science professor Arvind Narayanan stated that existing evidence suggests large companies are generally unwilling to allow their data to be used for model training. He noted that, despite consumer subscription plans often being 10 to 20 times cheaper, many enterprises still opt for enterprise versions billed per token, primarily due to data retention and enterprise IT governance requirements.
Meta previously faced resistance in acquiring training data. The company reportedly pushed a plan earlier this year to track employee computer usage, which sparked internal criticism and was paused in June. Meta did not respond to TechCrunch’s request for comment on the new pricing arrangement for Muse Spark.
Advanced models continue to engage in price wars.
This pricing also reflects how competition among cutting-edge model providers is shifting toward finer-grained cost structures. The day before, Anthropic released its new models, Fable and Mythos, and lowered the cost of processing cached tokens. OpenAI also significantly reduced prices for its latest models at the end of July.
As model capabilities converge, price, data acquisition methods, and enterprise clients' governance requirements are emerging as new competitive factors in the AI platform landscape.