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

Meta wants into the coding agent game, and it's doing what Meta does when it's playing catch-up: undercutting everyone on price. This week the company announced Muse Code, its first agent built for full software engineering tasks like planning changes, writing code, and validating results, running on a freshly updated model called Muse Spark 1.2. Meta is pitching pay-as-you-go pricing at $1.25 per million input tokens and $4.25 per million output tokens, positioning it below Anthropic's Claude Code and OpenAI's Codex.

But the real hook is the contributor tier, which Meta's chief AI officer Alexandr Wang told CNBC is more than 10 times cheaper than pay-as-you-go. The catch is that using it means opting in to help improve Meta's model. In other words, hand over your coding sessions and get a steep discount. That's a strategy Meta has used before with Muse Spark 1.1, its first paid model, released just weeks earlier as it tries to close ground on OpenAI and Anthropic with proprietary, aggressively priced tools rather than open-source releases.

Engineering leaders aren't biting. Ken Ringdahl, CTO of Emburse and a Snyk advisor, put it bluntly: there's no IP more important than source code, and he won't gamble it for a lower bill. Yunhao Jiao, CEO of TestSprite, goes further, pointing out that a coding session isn't just code — it's prompts, agent responses, and the corrections developers make along the way. His worry isn't necessarily what Meta does with that data today, but whether details could eventually surface in a way that helps a competitor replicate his team's work down the line. He admits there's no evidence for that happening, but it's enough to keep him away from the contributor tier.

There's also a practical wrinkle: token price isn't actually the metric that matters most, according to Michał Piszczek, CTO of Archdesk. He argues the number worth watching is cost per accepted change after review and repair, not the sticker price per token. Piszczek still sees a use case for Muse Code, just not on proprietary work — he'd point it at open source projects, personal work, or disposable prototypes, while reserving anything resembling a core product for tools offering zero-data retention.

Wang told CNBC that Meta is beginning to accept requests for zero-data retention arrangements, which would mean the company doesn't keep developer data for training. Nothing's been confirmed on timing, though. Until it is, the split Piszczek describes — cheap, data-sharing pricing for low-stakes work and stricter guarantees for anything that actually matters — looks like the realistic ceiling for Muse Code's adoption, at least among the people who spoke to The New Stack. Muse Code is out now in beta.