Meta offers a 21x discount, paid for with your code and conversations

After months of lagging, Meta shipped its coding agent Muse Code, powered by its new Muse Spark 1.2 model.

The agent is genuinely good. But that is not the story.

Performance first#

That puts it in the top ten, a position Meta has not held for a while.

Technically it runs sub-agents in the background with an event log protecting operations from crashes. So if something breaks mid-way through a long task, you do not start over. Anyone who has run an agent on an hour-long job knows that problem.

And now the offer#

Standard pricing is relatively low:

But Meta added a subscription tier 21 times cheaper.

On one condition: you let the company take your code and conversations to train future generations of its models.

When the offer is reasonable#

To be fair, it is not necessarily a bad deal:

A personal learning project. Your code carries no competitive value and the saving is real.

Code that is already open source. It is published; what are you protecting?

Experimentation. Before you have even decided whether to build something.

When accepting is a mistake#

Client code. Your contract almost certainly forbids it, and you probably have not read that clause in years.

Anything carrying data. Even in comments, even in variable names, even in a test file you forgot about.

A competitive advantage. If your code contains what distinguishes your company, you are handing it to someone who may build it for others.

A regulated entity. A bank, a government body, healthcare. The decision is not yours to make.

A question for you#

The offer is tempting, and I do not know whether users will take it.

Would you? Tell me, because I am asking seriously.

Common questions#

What is Muse Code?

Meta's coding agent powered by Muse Spark 1.2, running sub-agents with an event log that protects long tasks from crashes.

What does it cost?

$1.25 per million input tokens and $4.25 per million output, with a tier 21 times cheaper in exchange for allowing training on your code and conversations.

When is taking the offer reasonable?

On personal learning projects, already-open-source code, and during experimentation. Not for client code, anything carrying data, or a competitive advantage.

What gets forgotten in these offers?

That conversations are shared too, not only code. Explaining why you are building something and for whom hands over your strategy, and code can be rewritten while intent cannot be unshared.