Meta Just Put a Sticker Price on Your Codebase

Meta shipped its first coding agent on August 5, and the agent itself is the least interesting part. Muse Code, now in beta for macOS and Linux, is a terminal agent powered by the new Muse Spark 1.2 model. It plans changes, writes code, validates results, keeps async background agents alive for the whole session, and fans big jobs out to sub-agents in isolated worktrees. If that feature list sounds familiar, it should — it's a competent re-implementation of what Claude Code and Codex already do, backed by a model Meta says is "close to the best" on Terminal-Bench 2.1 and DeepSWE 1.1. Close to the best is not a reason to switch tools.

The pricing page is. Muse Spark 1.2 comes in two SKUs running the identical checkpoint with the same ~1M-token context window. The standard tier costs $1.25 per million input tokens and $4.25 per million output, and Meta commits to not training on your traffic. The "contributor" tier costs $0.10 in and $0.20 out — a 12.5x and 21x discount — in exchange for permission to use your prompts and completions to train future Meta models. Cached input drops from $0.15 to $0.002, a 75x gap, and cache hits are most of what an agent burning through a large repo actually bills.

Vendors have traded data for access before. Gemini's free API tier trains on your traffic; the paid tier doesn't. GitHub spent 2021 and 2022 fielding telemetry questions about Copilot. But those were binary gates: free-and-harvested or paid-and-private. Meta is the first to publish an explicit per-token exchange rate. The $1.15-per-million-input-tokens spread between tiers is, functionally, Meta's public bid for the right to train on your work. Nobody has priced that market before. Now there's a number on it, and every other lab gets to decide whether to outbid or ignore it.

"Prompts and completions" means your repo

Here's the part that should stop you before you swap API keys to save money. In a chat product, "prompts" means what you typed. In an agentic coding tool, the prompt is assembled by the agent — and it's stuffed with your source. Every file Muse Code reads into context, every diff it inspects, every test output and stack trace it feeds back to the model: all of it flows through as prompt tokens. Grant training rights on "prompts and completions" for an agent session and you've granted training rights on effectively every part of the codebase the agent touched. The comforting narrowness of that phrase is a chat-era intuition applied to a tool that reads your whole worktree.

That's why the who-should-use-this question is a legal question before it's an engineering one. Your own greenfield prototypes, open-source work where the code is public anyway, high-volume throwaway experimentation — the contributor tier is close to free and the trade is rational. Client code under an NDA, anything in a regulated sector, any repo where you can't personally answer "do we have the right to grant Meta a training license on this?" — the discount is irrelevant, because the permission isn't yours to sell at any price.

The discount is narrower than 21x

Even where the trade is legally clean, read the limits. The contributor tier is capped at 60 requests per minute per team, against 3,000 on standard — 2% of the throughput. Agentic workloads are exactly the wrong shape for that ceiling: lots of small tool-call round trips rather than a few fat requests, so you slam into the RPM wall long before the token quota matters. The cap is per team, not per key, so you can't shard around it. The contributor SKU also isn't on OpenRouter or other aggregators — direct Meta signup only — and add-ons like web-search grounding bill identically on both tiers, which drags a search-heavy workload's real blended discount down toward single digits. Meta engineered this tier so you can't quietly run production on it. It's for hobbyists and open-source at scale, which is to say: it's for training-data acquisition, dressed as pricing.

That reading fits Meta's position. Muse Spark is proprietary and closed-weight — a sharp break from the Llama era, and Zuckerberg's only comment on open-sourcing was "more to share soon." Meta is behind Anthropic and OpenAI in agentic coding, and the scarce input isn't compute, it's real-world traces of agents working in real repositories. A 21x discount is a data bounty. Meanwhile the standard tier undercuts Sonnet 5's $3/$15 while benchmarking in the neighborhood of frontier models, which is genuine price pressure even if you never touch the contributor tier.

What to actually do

If you're curious, the standard tier is the honest evaluation: no training on your traffic, competitive pricing, and Muse Code's design has a few genuinely good ideas — the append-only local event log of every model call, tool run, and edit is an auditability feature its competitors should copy. Zero-data-retention is only "beginning" to be available on request, per Meta's Alexandr Wang, so enterprises should treat that as a roadmap item, not a feature.

If you adopt the contributor tier for eligible work, enforce the boundary in infrastructure, not in developers' heads: separate API keys per tier, with CI or a proxy that pins client-facing and regulated repos to the standard key. "Remember to switch keys for client work" is a confidentiality breach with extra steps.

The larger shift is worth sitting with. For three years the industry treated training rights on developer data as something vendors took (or piously refused) — never something they priced. Meta just quoted the market: your agent traffic is worth about $1.15 per million input tokens. If that number strikes you as insultingly low for the contents of your repository, you now know exactly what to decline — and you've learned something about how the labs value the data everyone else has been handing over for a discount of zero.