News

Meta launches Muse Code, a coding agent priced to undercut rivals

Published Updated 6 min read

Read in : Français

Contents
  1. What exactly did Meta announce on August 5, 2026?
  2. How much does Muse Code cost, and what is the contributor tier?
  3. Is Muse Code as good as Claude Code and Codex?
  4. Why does this matter for vibe coders?
  5. What should you do this week?
  6. What we don't know yet
  7. Frequently asked questions
  8. Frequently asked questions
  9. Sources

Muse Code, Meta's coding agent, went into beta on August 5, 2026. It is a command-line tool in the same family as Claude Code and Codex, running on Meta's new Muse Spark 1.2 model. The real news is not technical: it is a very cheap "contributor" tier that you pay for with your data. Here is what Meta announced, what it means for you, and how to try it without hurting yourself.

What exactly did Meta announce on August 5, 2026?

Meta shipped two things on the same day: Muse Spark 1.2, a coding-optimized model, and an "early beta" of Muse Code, the agent built to run it. The announcement appeared on Meta's developer blog and the AI at Meta research blog, and was covered by TechCrunch, VentureBeat and Engadget.

What Meta's own documentation describes:

Mark Zuckerberg summed up the pitch on X, as quoted by VentureBeat: "Install Muse Code with one line and you can start on our contributor tier."

How much does Muse Code cost, and what is the contributor tier?

Muse Code is billed by model usage, with two tiers that serve the same Muse Spark 1.2 model.

TierInput (1M tokens)Cached inputOutput (1M tokens)Trains on your data
Standard (muse-spark-1.2)$1.25$0.15$4.25No, per Meta
Contributor (muse-spark-1.2-contributor)$0.10$0.002$0.20Yes

Standard prices are listed on Meta's developer blog. Contributor prices are not spelled out on that page: VentureBeat and Engadget (citing The Wall Street Journal) report the same figures. The Next Web published a different number ($0.30) that no other outlet matches, so go with the consistent VentureBeat and Engadget figures and check your Meta console before you budget.

Three details matter:

Is Muse Code as good as Claude Code and Codex?

On Meta's own numbers, Muse Code is close to the leaders but not ahead. The charts in the official post are images; the figures below come from VentureBeat, which reported them. These are vendor benchmarks, not independent tests.

Benchmark (Meta's figures)Muse Spark 1.2 in Muse CodeBest rival cited
Terminal-Bench 2.182.9%Claude Opus 5 in Claude Code: 86.7%
DeepSWE 1.159.3%Claude Opus 5: 65.0%
Meta internal coding bench70.6%Claude Opus 5: 79.4%

VentureBeat also notes that on Terminal-Bench 2.1, Muse Spark 1.2 edges out GPT-5.6 Terra in Codex (81.8%). Put simply: one step below Claude Opus 5, level with Codex, and far cheaper on the contributor tier. To understand what these tests measure, read our guide to picking an AI model for coding.

Why does this matter for vibe coders?

Muse Code lowers the floor price of a terminal coding agent. Until now, cost was the main thing stopping vibe coders from moving from an all-in-one builder to an agent: a long session on a premium model adds up fast. At a few cents per million tokens, experimenting is close to free.

The trade-off is simple: you pay with your code. For an open-source side project or a throwaway prototype, that is often fine. For a client app, an internal company tool, or anything holding personal data, it is not. Client code is not always yours to share, and sending it to a third party for training may breach your contract. Our piece on who owns AI-generated code covers the rights questions.

Also, Muse Code is a terminal tool. If you are coming from Lovable or Bolt, the learning curve is the same as with Claude Code, and our Claude Code beginner's guide explains the workflow, which carries over here.

What should you do this week?

If you want to try it, do it on neutral ground. Here is the process.

  1. Pick a low-stakes project: a personal prototype, a public repo, a practice exercise. No client code, no .env file with real keys.
  2. Check which tier is active: the model name (muse-spark-1.2 or muse-spark-1.2-contributor) decides whether Meta trains on your data. Confirm it before the first task.
  3. Move secrets out of the folder: an agent reads whatever is lying around. Our article on why API keys must never live in your front end applies just as much to what you send a model.
  4. Set a budget: reasoning tokens count as output. Run one task, check usage, then extrapolate. Our breakdown of what a vibe-coded app really costs shows how.
  5. Compare on the same task: give Muse Code and your usual agent the same prompt, then compare the diff, the tests and the time spent fixing things.

A starter prompt for a clean test:

Read this project without changing anything. Describe its architecture
in 10 lines, list every file that contains secrets or personal data,
then propose a 5-step plan to add [feature].
Wait for my approval before writing any code.

What we don't know yet

Several points in the August 5, 2026 announcement are still vague. The list of countries eligible for the contributor tier is not published. Windows support is unclear: the developer blog shows a Windows tab, while the research blog mentions only macOS and Linux. And Muse Code is a beta, so expect commands and limits to change. To place Muse Code among other tools, our comparison of vibe coding tools and our Cursor vs Claude Code vs Codex comparison are the right starting points.

Frequently asked questions

Frequently asked questions

Is Muse Code free?

No. Installing Muse Code is free, but Muse Spark 1.2 usage is billed per token. The standard tier is listed at $1.25 per million input tokens and $4.25 per million output tokens. The contributor tier, reported at $0.10 and $0.20, is far cheaper, but Meta can train on your prompts and completions, and usage is capped over a rolling five-hour window.

Does Meta train on my code when I use Muse Code?

It depends on the tier. On the contributor tier, yes: Meta says data "may be used to improve our products". On the standard tier, VentureBeat reports that Meta says it does not train on prompts or completions. The choice is made through the model name, so check it before any sensitive project.

Can Muse Code replace Claude Code?

For a personal prototype, it can do the job for less money. On Meta's published benchmarks, Muse Spark 1.2 still trails Claude Opus 5 in Claude Code (82.9% vs 86.7% on Terminal-Bench 2.1). Those are vendor numbers: the only test that counts is yours, on your project, with the same prompt given to both agents.

Can I use the contributor tier on client work?

It is a bad idea. Sending a client's code to a third party for training may breach your contract or an NDA, and it raises GDPR issues if the code contains personal data. Use the standard tier, or ask Meta for zero data retention, and get your client to approve the tool.

Sources