Meta has started charging two different prices for the same AI model. The gap between them tells you a lot about what AI companies actually need right now: not more computing power, but more real examples of how people use these tools.
The model is called Muse Spark, and it powers Meta's new coding assistant. Under a standard plan, a million input tokens, the chunks of text the AI reads, cost $1.25, and a million output tokens cost $4.25.
Agree to let Meta use your prompts and results to train future versions, and those same amounts drop to 10 cents and 20 cents. That is a price cut of roughly 95 percent for giving up your privacy.
Muse Spark itself is a bigger deal than it looks. It is Meta's first closed model, meaning outside developers cannot download and inspect it the way they could with Meta's earlier Llama models.
That shift came after Alexandr Wang, who built and sold the data company Scale AI, joined Meta last year to fix a stalled AI effort. Ending the open approach and building a paid coding product were both part of that fix.
Getting good training data for coding agents is hard, because most of it lives inside private company laptops, not on the open internet. Meta already learned this the hard way.
Earlier this year it launched a program to record its own employees' keystrokes, mouse movements, and screens to use as training data. Employees revolted, a petition against the program gathered close to two thousand signatures, and Meta paused it after a security failure exposed some of that sensitive data to the wrong people inside the company.
Buying data from paying customers, instead of pulling it from unwilling staff, is Meta's next attempt to solve the same problem.
Here is the part worth remembering. Careful, well resourced companies do not take these deals.
Research on enterprise AI use shows large companies keep paying full price for private, usage based plans instead of switching to subscription plans that cost ten to twenty times less. They do this specifically because the private plans do not train on their data.
That gap in behavior tells you these companies view their prompts and files as worth protecting, even at a steep price.
This is not only a Meta story. Anthropic and OpenAI have both cut prices sharply this year too, and the overall cost of running AI models has fallen to its lowest point yet.
But cheaper tokens do not always mean a cheaper bill, since agentic tools use far more tokens per task than simple chat does. Watch for this same trade, a steep discount in exchange for training rights, to show up in AI tools built for sales, legal, or customer service work, not just coding.
Before taking any AI discount tied to data sharing, know exactly what you are handing over. Once your workflow shapes a public model, you cannot take it back.