[Meta](#/company/meta) launched its first dedicated coding agent, Muse Code, in beta for macOS and Linux around August 5 — a terminal-based tool built to plan, write, and validate multi-step software changes, aimed squarely at the market Anthropic's Claude Code and OpenAI's Codex currently split. It runs on Muse Spark 1.2, an updated version of Meta's coding-tuned foundation model, available through the Meta Model API. Unlike a chat-based assistant, Muse Code keeps subagents alive for an entire session, letting them gather information in the background and report back rather than resetting context between turns.
The product itself reads as a reasonable, late entrant to a crowded field. What makes it a story is the price — and what the price actually buys.
Two tiers, one very large gap
Muse Code ships with two pricing tiers on the same underlying model. The Standard tier runs `muse-spark-1.2` at $1.25 per million input tokens and $4.25 per million output tokens — in the same range as competing coding-agent pricing. The Contributor tier runs `muse-spark-1.2-contributor` at $0.10 per million input tokens and $0.20 per million output — roughly 12 times cheaper on input and more than 20 times cheaper on output, per Meta's own pricing page. Cached input — repeated context within a session — is billed separately: $0.15 per million on Standard versus $0.002 per million on Contributor.
Muse Code: Standard vs. Contributor
| Standard muse-spark-1.2 | Contributor muse-spark-1.2-contributor | |
|---|---|---|
| Input tokens (per 1M) | $1.25 | $0.10 |
| Output tokens (per 1M) | $4.25 | $0.20 |
| Cached input tokens (per 1M) | $0.15 | $0.002 |
| Rate limit | 3,000 req/min, 4M tokens/min | 60 req/min, 2.1M tokens/min |
| Data use | Never used to improve Meta's products, per Meta | May be used to improve Meta's products, per Meta |
Both tiers carry the same 1M-token context window and full agent capabilities — Meta isn't holding back functionality to push the discount tier, and a user can switch tiers mid-session with a `/model` command. The entire difference is what happens to the data afterward.
What "may be used to improve Meta's products" actually means
Meta's pricing page states the trade in plain terms: data from the Contributor model "may be used to improve Meta's products," while the Standard model's data is "never used to improve Meta's products." Meta says Contributor-tier data is anonymized and aggregated before use. What Meta has not published is the anonymization methodology itself, or an independent audit confirming it holds up against re-identification — a gap several outlets covering the launch flagged without Meta directly answering it on the record.
The harder problem isn't anonymization at intake. It's what happens after the data is used for training. Once code passes into a model's weights during training, there is no extraction procedure that pulls a specific contributor's snippets back out — a trained model doesn't store a retrievable copy of any single input, it stores adjusted weights shaped by millions of them. Privacy researchers covering the launch have pointed out that this makes a request under a law like the EU's GDPR — Article 17, the "right to erasure" — effectively unsatisfiable once training has happened, regardless of Meta's intentions or a user's later regret. Meta has not disputed that description; it has simply not addressed it directly.
Where this sits against Muse Glimmer
Muse Code's launch lands in the same stretch as [Muse Glimmer](#/article/meta-muse-glimmer-30b-open-agent-model), Meta's open-weight 30-billion-parameter agent model — a separate release, aimed at a separate audience. Glimmer is a model anyone can download and run locally, with no data ever leaving the user's own hardware. Muse Code's Contributor tier is the opposite design: cloud-hosted, cheap, and dependent on the data flowing back to Meta. Together they read less like a single strategy than like Meta testing two different bets on how to catch up in agentic coding — one open and local, one closed and data-funded — in the same two-week window.
Anthropic's Claude Code and OpenAI's Codex don't currently offer an equivalent discount-for-data tier at this scale; both vendors' standard pricing is closer to Muse Code's Standard rate than its Contributor rate. Whether that changes is the next thing worth watching — a price this aggressive puts pressure on competitors to either match it or defend why they haven't.
- Meta launched Muse Code, an agentic coding tool for macOS and Linux, around August 5.
- Contributor-tier pricing is $0.10 per million input tokens and $0.20 output, versus Standard's $1.25 and $4.25.
- On Contributor, Meta may use prompts and completions to improve its products; Standard data never is.
- Contributor is capped at 60 requests and 2.1M tokens per minute, versus Standard's 3,000 and 4M.
- Once code trains a model there's no extraction — an erasure request becomes effectively unsatisfiable, researchers say.