Meta is at it again trying to keep up with AI behemoths like Anthropic and OpenAI. With its newest model, Muse Code, Meta’s new pitch is to provide developers a cheaper coding agent option compared to Claude Code or Codex.
The catch? If you opt-in, Meta trains on all your coding data.
The CTO of Emburse was quoted in The New Stack, “Security, data privacy, [and] data security is paramount, and it’s not something we will trade off for a lower-cost option.”
But as companies still try to rein in their token spending, is it worth the trade?
I asked my leadership team what they thought; here’s what they said.
Why Should We Trust A Company Known For Privacy Violations?
Aby Jacob (VP of Engineering)
Meta has shipped a genuinely usable agent at a price point nobody else will provide, so credit where it is due. But let's be honest about what the contributor tier actually is. We've all spent the last two years keeping Chinese open weight models out of our stacks because of data exfiltration fears. Nobody ever proved Qwen or DeepSeek were sending data home, but no CISO would ever take that gamble.
Well here is Meta, and they're not even asking us to suspect anything. The data collection is printed right on the pricing page: "Opt in to help improve the model" is a polite way of saying "Hand us the most valuable thing your engineering team produces." And it's not just source code. It's your prompts, the agent's plans, the failed attempts, and the corrections your senior engineers make before accepting anything.
So if we blocked the Chinese models over a possibility, I'm not sure how Muse Code's contributor tier gets a pass when they clearly tell us they’re going to use our data.
And that's because this is a go-to-market play, not a technical breakthrough. OpenAI and Google got to this market early and own the enterprise mindshare, so Meta is playing catch-up, and the only lever left for a late entrant is price. Cheap tokens will absolutely win the price-sensitive crowd.
“So if we blocked the Chinese models over a possibility, I'm not sure how Muse Code's contributor tier gets a pass when they clearly tell us they’re going to use our data.”
But keeping those customers means the model has to actually be good, and ensuring quality means needing training data at a scale you simply can't scrape anymore (the public code corpus is picked clean, everyone's already eaten from that buffet).
So the discount funds the data pipeline, and the data pipeline funds the model that retains the customers the discount brought in. Honestly, it's a beautiful flywheel from a marketing/business point of view. Meta is trying to win either way: you pay full price, or you pay in the training data they need to catch up. The breakthrough is in the pricing table, not the model card or new innovation.
Everyone’s Fears Are Misplaced, But I Still Won’t Use Muse Code
Nikhil Gopinath Kurup (SVP of Engineering, ML)
The fear that Meta's model remembers your team's work and later hands it to a competitor building the same feature is a misconception. Weights aren't a retrieval index; verbatim memorization does happen, but it concentrates on text that appears thousands of times in a corpus, license headers, and the boilerplate everybody copies.
One session's business logic seen once in a training run isn't going to show in a future LLM response just because someone at another company asked a vaguely similar question.
The things that should actually worry you are duller: retention windows, breach, subpoena, and the contractors reading flagged sessions in the annotation queue. And that’s where the trust question shows up, and it's more relevant because every item on that list is governed by policy rather than by math/models/weights.
"While the engineering fear is misplaced, my answer is still no."
Meta’s Chief AI Officer, Alexandr Wang, says Meta is "starting to accept requests" for zero-data retention, and the article notes Meta hasn't confirmed if or when that actually ships.
This is a company that torrented 81.7TB of pirated books to train Llama (a judge let the training through on fair use; publishers are now suing again), and paid the FTC $5 billion over how it handled user data. So a significant amount of trust needs to be built to believe "we won't retain this."
So, while the engineering fear is misplaced, my answer is still no.
Which brings me to the other point that Meta's standing with developers has always been strange. Cassandra, React, and Prophet were real gifts, and Llama was the open weight release that showed a lot of people what was possible outside a frontier lab.
Good will from these projects are good, but does it translate into "Type your production repo into our terminal"?
I also can't tell what the goal is. Is it the best coding model, most-used coding model, cheapest coding model, or a data program with a product wrapped around it?
The contributor tier reads like the latter, while $1.25 per million input tokens and $4.25 per million output tokens reads like Spolsky's commoditize-your-complement move: If you can't be the frontier, you make frontier-ish coding cheap enough that nobody else can afford to fund being the frontier either.
Meta did roughly that with Llama from behind, and it reset what people expected to be charged.
Honestly, it may not matter. Prices are falling anyway with the Chinese labs shipping and OpenAI and xAI cutting, so the price war is coming, and everyone I know is watching their token bill in a way they weren't last year.
Another cheap agent is welcome. The contributor tier is easy enough to not check the box on.
Nobody in This Industry Will Trade Their Code for Cheaper Tokens
Hari Chandrasekhar (SVP of Engineering, Core)
We hit this concern about data privacy and security well before anyone put a price on it.
When we started evaluating coding agents at Sedai, the first question was not “What are the cost or benchmark scores?”, but rather, “What do these systems retain and what do they train on?”
Talking to peers, privacy is consistently the top objection, regardless of company size. A two person startup with a differentiated algorithm has the same risk as an enterprise; this is simply a non-starter across the board.
“Given Meta's track record on privacy, I still don't see this becoming a durable pattern in the developer community.”
On Meta's position: they are about 17 months behind Claude Code's first release and 14 behind its GA, and Muse Code is still in beta. In this market, that’s practically decades, so it’s understandable that they have to trade something to get in the door. But they chose to trade price for training data, something I don’t think anyone working in this industry will do.
The pricing model itself is not new, and it’s the same bargain Meta and Google have run for two decades: cheap or free service in exchange for data that improves targeting. That held because most users accepted the trade, knowingly or not, and because the custodians kept the data. Google still states publicly that they do not sell personal information to anyone.
But coding agents break that equilibrium in two ways.
First, the audience. Developers and engineering managers understand the trade well enough to price it, and based on my past experience, most will just pay the higher tier.
Second, the failure mode is different. Using data to target users with ads is incredibly low-risk. But with training, your proprietary logic can plausibly surface in someone else's completion. To be fair, there is no evidence of that happening. But "no evidence yet" is a weak assurance when the mechanism is entirely plausible.
So where does Muse actually fit? Open source, side projects, throwaway prototypes. It absolutely should not touch product code at any company size.
My read is that Meta priced this to get attention and installs, not because they expect the contributor tier to carry the business. The counterpoint is that early user reports say the contributor tier is the default after install, which suggests they do want volume. Given Meta's track record on privacy, I still don't see this becoming a durable pattern in the developer community.
Cheaper tokens shouldn't cost you your code. Sedai routes each request to the most efficient model for the task. See how.