Meta offers up to 95% off to those who share their usage data with its Muse Spark model

Meta has introduced an unusual pricing scheme for Muse Spark, its new model designed to power coding agents and other kinds of agents: those who agree to let their prompts and the model's responses be used to train future versions pay, on average, around 95% less than those who do not. According to the figures cited in the article, under the standard agreement one million input tokens costs $1.25, while under the "contributor" option that same million costs just 10 cents. For output tokens, the standard price is $4.25 per million, compared with barely 20 cents under the contribution scheme.

The move makes more sense in light of the difficulties Meta has had obtaining quality training data. The article recalls that earlier this year the company launched an initiative to monitor how its own employees used their computers, a measure that drew strong internal criticism and was suspended in June. Meta did not respond to TechCrunch's questions about this new pricing policy, leaving it unclear to what extent the discount is intended to directly offset that data shortfall.

The value of this kind of information for improving agents is illustrated by a quote from Mario Zechner, developer of the open-source harness Pi, who told TechCrunch that the leap in capabilities seen in coding agents between April and October 2025 was due to Claude Code storing user sessions by default and using them for reinforcement training. The problem, the article notes, is that as model makers try to take their agents beyond programming into other professional workflows, the lack of clear digital traces in those tasks makes them harder to evaluate and improve.

Against that, Arvind Narayanan, a computer science professor at Princeton, adds a relevant nuance: large companies already show through their behavior that they prefer not to hand over their data for training. As the article quotes, Narayanan noted on social media that these companies stick to enterprise plans billed per token even though consumer subscriptions such as Claude Max or ChatGPT Pro work out 10 to 20 times cheaper or even more; the main difference between the two types of plan, he points out, lies in data retention and enterprise IT governance. On that reading, Meta's scheme would be an explicit acknowledgment of that dynamic: rather than assuming customers will hand over their data, the company is offering direct financial compensation in exchange. Meta's own pricing material describes the "contributor" route as a way to "lower the barrier to entry for prototyping, testing integrations and scaling experiments where training on customer data is acceptable".

Narayanan further suggests that this kind of financial incentive could push large companies to be more careful in distinguishing which data is genuinely proprietary and which could be shared with model providers.

The article ultimately frames this move within a broader price competition among the major AI labs. It notes that Anthropic's new Fable and Mythos models, unveiled the day before the article was published, came with reduced costs for processing cached tokens, and that OpenAI's most recent models received significant price cuts in late July. In that context, Meta's model would add a different variable to a simple price cut: it is not just about charging less, but about explicitly tying price to access to the customer's usage data, something no other competitor mentioned in the article appears to have formalized in the same way.