Meta Muse Code: the cheap tier is paid for with your code

Meta shipped a terminal coding agent today. Muse Code installs with one curl command, runs on macOS and Linux, and points Meta's Muse Spark 1.2 model at your whole repository. It is a direct shot at Claude Code and Codex, and the interesting part is not the agent. It is the price sheet. Meta will sell you tokens at roughly a tenth of list — if you let it train on what you feed the agent.

What actually happened

Per Meta's research blog, Muse Code is in beta and installs via curl -fsSL https://dev.meta.ai/install.sh | bash. It takes on "complete software engineering tasks across large repositories: planning changes, writing code, and validating the results." Three design choices stand out:

• Async background agents that stay alive across a session, backed by a local event log that makes runs replay-exact and restart-safe.
• Parallel sub-agents in isolated git worktrees. TechCrunch quotes Meta: "Your working copy is never touched. In testing we had it build six features for a game simultaneously with no collisions."
• Bundled skills — /plan, /grill, /goal — shipped as first-class commands rather than prompt files you maintain.

Muse Spark 1.2 was co-trained with the agent, carries a 1M-token context window, and is listed on OpenRouter at $1.25 input / $4.25 output per million tokens — the same rate as 1.1.

Then there is the second tier. On the Meta Model API, Unite.AI reports a contributor tier at $0.10 input / $0.20 output per million, with cached input at $0.002. Standard-tier prompts are not used to improve Meta's products. Contributor-tier prompts are. That is the trade, stated plainly.

Meta published benchmark charts for Terminal-Bench, DeepSWE, and an internal coding bench — as images, with no methodology write-up. Treat them as marketing until someone reproduces them on your code.

Why AI coding agent pricing matters for your business

A 90% discount on tokens is a licensing decision, not a procurement one. The contributor tier is cheap because your repository becomes training data. For a solo project, fine. For client work under an NDA, for anything with a customer list in the fixtures, for code you intend to sell — that is a term your contracts probably already prohibit, and nobody is going to catch it in a code review. Decide the tier at the org level and enforce it in config, not in a Slack message.

Cost per task beats cost per token. A cheaper model that needs three passes is not cheaper. Before you switch a team over, run the same five real tickets through both agents and compare total spend and human review time. We have watched the "cheap" option lose that test more than once.

The agent is portable; the harness is not. Background agents, worktree fan-out, and bundled skills are genuinely good engineering, and they are also Meta-specific. Keep your prompts, evals, and CI glue in your own repo so the next agent that undercuts this one is a swap, not a migration.

Sources: Meta AI Research, TechCrunch, OpenRouter.