Meta Will Pay You 95% Off to Let It Train on Your Prompts

Meta's Muse Spark model offers a huge discount in exchange for your prompts and outputs. The training-data trade just got explicit.

Published: 2026-09-04 Category: Quick Take Sources: TechCrunch

The Story

Most AI tools let you opt out of sharing your usage with the model provider to improve future versions. Meta has taken that idea and put a price tag on it. For its new Muse Spark model, aimed at powering coding and other agents, Meta is offering contributors an explicit discount averaging about 95% in exchange for sharing their prompts and model outputs for training future models.

The numbers are dramatic. Under a standard agreement, 1 million input tokens cost $1.25; under the "contributor" pricing model, they cost just 10 cents. For output tokens, the standard rate is $4.25 per million — but the same million costs just 20 cents for contributors. In both cases, that's roughly a 95% cut.

Meta has had a rough time gathering training data through other means. An initiative to track the computer usage of its own employees, launched earlier this year, attracted wide internal criticism and was paused in June. Meta did not respond to TechCrunch's question about the new pricing model.

Why It Matters

This kind of user data is increasingly vital to making agentic tools work well. Mario Zechner, the developer behind the open source harness Pi, noted that a big jump in coding-agent capabilities between April 2025 and October 2025 came largely because Claude Code, by default, stored all coding sessions and used them for reinforcement learning. But as model builders push agentic tools into fields beyond software engineering, their ability to evaluate and improve those tools is blocked by the complexity and lack of digital traces in many professional workflows.

Princeton professor Arvind Narayanan observed that there's good evidence large companies don't want their data used for model training — they stick with token-billed enterprise plans even though consumer subscription plans like Claude Max and ChatGPT Pro are discounted 10x-20x or more, precisely because the enterprise plans differ on data retention and IT governance.

The Takeaway

The explicit nature of Meta's offer is the news here. Instead of burying data-sharing in terms of service, Meta is pricing it out in the open: your data has a discount value, and here's the deal. Narayanan suggested this could actually incentivize large companies to be more diligent about which data is truly proprietary and which could safely be shared with model providers.

It also feeds directly into the price war between frontier labs. Anthropic's newest Fable and Mythos models, released this week, came with lower costs for cached tokens, while OpenAI's latest models saw major price cuts at the end of July. When your usage data can purchase a 95% discount, the cost of intelligence and the value of your data have become the same negotiation.

Source: TechCrunch, "Meta is paying to peek at how you use their latest AI model" (Sept 3, 2026).