Meta’s new coding agent is cheap (but it’ll cost you your data).

Meta wants into the coding agent business, and it's trying to buy its way in with a price tag rather than raw capability. This week the company unveiled Muse Code, a terminal-based agent built on the newly updated Muse Spark 1.2 model, aimed at the kind of end-to-end work developers now expect from tools like Claude Code: planning changes, writing the code, checking that it actually works. On paper, the pay-as-you-go pricing of $1.25 per million input tokens and $4.25 per million output tokens already undercuts OpenAI and Anthropic. But the real hook is the contributor tier, which Meta's chief AI officer Alexandr Wang says runs more than ten times cheaper than that pay-as-you-go rate.

The discount comes with strings attached, and they're the kind that make engineering leaders nervous. To get contributor pricing, developers have to opt in to feeding their coding sessions back to Meta to improve its models. Ken Ringdahl, CTO of Emburse and an advisor to Snyk, put it bluntly: there's no intellectual property more valuable than source code itself, and he's not gambling with it for a discount. Yunhao Jiao of TestSprite goes further, pointing out that a coding session isn't just code — it's prompts, corrections, agent reasoning, the whole trail of how a team actually builds something. His worry isn't that Meta misuses that data directly, but that a model trained on it could later help a competitor replicate the same feature. No proof that happens yet, he admits, but enough doubt to stay away.

This is part of a broader pivot for Meta, which has spent most of this year edging away from its open-source identity. Muse Image landed in early July, Muse Spark 1.1 followed as Meta's first paid model, and now Muse Code extends that same undercut-the-incumbents strategy into agentic coding. Zuckerberg's company is clearly betting that being dramatically cheaper than Anthropic and OpenAI will pull in enough users to close the gap, even if the underlying model isn't obviously better.

The trouble is that price isn't actually the constraint most companies are wrestling with. Jiao says his team doesn't even view coding agent spend as a meaningful expense, data-sharing risk aside. And Michał Piszczek of Archdesk argues that token price is the wrong yardstick entirely — what matters is cost per accepted change after review and repair, a number that has nothing to do with contributor discounts. Piszczek does see a plausible niche for Muse Code: open-source repos, personal projects, throwaway prototypes, anything where leaking training data isn't a business risk. For core product code, he wants zero-data retention, something Wang says Meta is starting to entertain but hasn't committed to on any timeline.

So Muse Code launches into beta with a strange split personality — genuinely cheap, genuinely capable by Meta's account, but structurally unattractive to the exact enterprise customers who'd generate the most usage. Until Meta nails down real zero-retention terms, the contributor tier looks less like a pricing innovation and more like a data-collection scheme wearing a discount as a disguise.