Meta Muse Code's Contributor Tier: 21x Cheaper Tokens If You Hand Over Your Code

Meta shipped its first coding agent, Muse Code, and did something no major lab had done so openly: it put a price on your source code. Opt into training rights and your output tokens drop from $4.25 to $0.20 per million — a 21x discount. Decline, and you pay standard rates.

Muse Code itself is free; you pay for the model underneath it, muse-spark-1.2. The standard tier is $1.25 input / $0.15 cached / $4.25 output per million tokens. The contributor tier — muse-spark-1.2-contributor — is reportedly $0.10 input / $0.002 cached / $0.20 output. In exchange, Meta gets the right to train on the code you send through it.

The Raw Economics

The discount is real and large. Model a heavy agentic coding day: 20M input tokens and 4M output tokens.

• Standard tier: 20M × $1.25/M + 4M × $4.25/M = $25.00 + $17.00 = $42.00/day
• Contributor tier: 20M × $0.10/M + 4M × $0.20/M = $2.00 + $0.80 = $2.80/day

That's a 15x total reduction on this mix, and it widens toward 21x as output-heavy work dominates. Over a year of daily use, the difference is roughly $14,300. For a bootstrapped developer or a student, that is not a rounding error — it's the difference between affordable and not.

What You're Actually Trading

The discount isn't charity — it's a data acquisition strategy. Your prompts, your codebase context, and the model's outputs on your code become training material. That has three concrete implications:

• Confidentiality. Proprietary business logic, secrets accidentally left in context, and unreleased features could all end up shaping a model that competitors also use. Even with de-identification, code is often self-identifying.
• Licensing. If you work under an employment contract or client agreement with IP or confidentiality clauses, you may not have the right to grant Meta training rights on that code at all. This is a legal question, not just a pricing one.
• Regurgitation risk. Models can memorize and reproduce training data. Code you contribute could, in principle, resurface in another user's completion.

A Simple Decision Framework

Take the contributor tier when: you're working on open-source code that's already public, personal projects with no confidentiality obligations, throwaway prototypes, or learning exercises. If the code is already on a public GitHub repo, you've given up little by letting Meta train on it — and the 21x discount is pure upside.

Pay standard rates when: the code is proprietary, client-owned, or covered by an NDA; you handle secrets or PII in context; or you simply can't audit what leaves your machine. The premium is the cost of keeping your code out of a training set — and for commercial work, that's usually cheap insurance.

The Precedent That Matters

Muse Code makes explicit what has been implicit across the industry: your data has a dollar value, and providers are now willing to name it. Expect more tiered "data-for-discount" pricing to follow. The healthy response isn't reflexive rejection — it's treating your code like any other asset with a price, and deciding case by case whether the discount is worth what you're giving up. For public code, take the deal. For anything you'd be uncomfortable seeing in a competitor's autocomplete, don't.

Compare Muse Code's tiers against other coding models for your workload in our AI Cost Calculator.