Meta's Muse Code Costs 12x Less If You Hand Over Data

Meta released Muse Code on August 5, a terminal coding agent running on a new model called Muse Spark 1.2. The agent is competent. The price list is the story. One model, two rates: $1.25 per million input tokens on the standard tier, or $0.10 per million on what Meta calls the contributor tier. Nothing about the model changes between them. The only difference is that contributor-tier users grant Meta permission to train future models on their prompts and completions.

That is a 12.5x discount on input and a 21x discount on output, paid for in source code. Meta is not competing on margin here. It is buying a corpus.

RelatedMeta's Muse Spark 1.1 Chases Anthropic and OpenAI

• Install is one line. curl -fsS https://dev.meta.ai/install.sh | bash, macOS and Linux, terminal only, beta.
• Standard tier: $1.25 / $4.25 per million input / output tokens, $0.15 cached input. Meta commits that this traffic is not used for training.
• Contributor tier: $0.10 / $0.20 per million, $0.002 cached input, in exchange for training rights over everything you send and everything it returns.
• Independently measured at 86.6% on SWE-bench Verified by vals.ai, which ties it exactly with Grok 4.5 and puts it ninth on our board.

What does Muse Code actually do differently?

It fans out. Give it a job large enough and it splits the work across background sub-agents, each running in its own isolated git worktree, so the implementation never touches your working copy while reviewers run alongside in the background. Mark Zuckerberg's framing at launch was that when a job is big enough, it separates into sub-agents working in parallel. Meta's demo built six features for a game simultaneously without collisions.

Worktree isolation is the honest engineering choice in that design. Most terminal agents that claim parallelism either serialize under the hood or let concurrent edits stomp each other in a single checkout. Giving each sub-agent a real, separate checkout costs disk and setup time and buys you the one property that matters: a failed sub-agent cannot corrupt the others or your uncommitted work. Anthropic's Claude Code and OpenAI's Codex both support parallel work, but Muse Code makes the isolation the default rather than something you opt into.

Is Muse Spark 1.2 good enough to matter?

Yes, with a ceiling. vals.ai, which runs every model itself on one neutral bash-only harness rather than accepting vendor numbers, scores Muse Spark 1.2 at 86.6% on SWE-bench Verified. That is a 4.6-point jump over Muse Spark 1.1 in under a month, and it lands exactly level with Grok 4.5. It sits about 10 points below Claude Opus 5 at 97.0%.

The weakness shows up in the difficulty breakdown. On tasks vals.ai classes as one to four hours of human work, Muse Spark 1.2 resolves 64%, against 90% for Claude Opus 5 and 74% for Claude Opus 4.8. On tasks under 15 minutes it hits 92%, close to the leaders. Read that as a model that is excellent at bounded, well-specified changes and still loses the thread on long agentic runs, which is exactly the workload the parallel sub-agent design is trying to paper over.

Meta's own numbers, which use Meta's own scaffolding, put Muse Spark 1.2 at 82.9% on Terminal-Bench 2.1 against 86.7% for Claude Opus 5, and 59.3% on DeepSWE 1.1 against 65.0%. Notably, Meta published a comparison where it loses. That is unusual and worth crediting.

How does it compare to Claude Code and Codex?

What it means for the market

The signal for investors is not that Meta built a better coding agent, because on the neutral harness it did not. It is that Meta is willing to price inference at roughly a rounding error to acquire proprietary code and the diffs that fix it. Anthropic and OpenAI sell coding tokens as a product with a margin attached. Meta, which has no such revenue line to protect, can treat the same tokens as customer acquisition cost for training data.

Watch two things. First, whether Anthropic or OpenAI answer with a data-for-discount tier of their own. Both have so far treated "we do not train on your API traffic" as a paid-tier guarantee and a selling point to enterprises, and a price war fought in data rights would put that positioning under real pressure. Second, whether large enterprise buyers treat contributor-tier pricing as an outright procurement disqualifier. If they do, the discount only ever reaches indie developers and students, which is a much smaller and much less valuable corpus than Meta presumably wants.

Microsoft is the exposed name to watch here. GitHub Copilot's per-seat subscription assumes token costs stay meaningful; a credible competitor at $0.10 per million input tokens changes what a seat is worth, even if the underlying model is ten points weaker. None of this is investment advice, and a beta product with no GA date is a weak basis for anything. But the pricing structure is a genuine strategic move, not a promotion.

RelatedDeepSeek V4 Flash claims 82.7 on Terminal-Bench 2.1

Who should refuse the contributor tier outright?

Anyone whose code is not theirs to license. That covers client work under NDA, contractor engagements, anything under a customer data processing agreement, regulated codebases in health or finance, and any employer whose IP policy you have not actually read. The consent is not narrow: it is prompts and completions, which means the file contents you paste in as context are inside the grant, not just the questions you ask.

For a personal side project or open-source work you already publish, the calculus is different and the discount is real money. Just make the decision deliberately once, at install time, rather than discovering it in a billing dashboard three months in.

Our take

Muse Code is a good agent at a fair price and a very good agent at a price that should make you read the terms twice. The engineering is genuinely thoughtful, particularly the worktree isolation, and a 4.6-point independent gain in a month is a real result rather than a benchmark stunt. But the two-tier structure is the actual product decision, and it deserves to be discussed as one rather than buried in a pricing page.

There is a version of this that is fine: developers who understand the trade, make it knowingly, and get frontier-adjacent coding help for almost nothing. There is another version where a junior engineer runs a curl-pipe-bash installer, picks the cheap option because it is the cheap option, and quietly licenses their employer's codebase to a company that competes with their employer. Meta has not made it hard to end up in the second version. That is the part worth pushing back on.

Original analysis by GenZTech. Reporting on the launch from TechCrunch; benchmark figures measured independently by vals.ai.