Muse Code AI is Meta’s terminal-based coding agent. It launched in beta on August 5, 2026, and left beta on August 31, 2026. It plans changes, edits files, runs commands and tests, and validates results across large codebases. It runs on macOS and Linux and uses Meta’s Muse Spark and Muse Wonder models. You can pay per token ($1.25 in / $4.25 out per million on the standard tier) or subscribe from $5 a month. A much cheaper “contributor” tier exists, but Meta may train on your code.
Most Muse Code articles are launch news. This one answers the question that matters once the hype fades: what will it actually cost you, and what do you give up for the lowest price?
Muse Code at a glance
| Detail | Information |
| Developer | Meta Superintelligence Labs (Meta Platforms) |
| Type | Terminal (command-line) AI coding agent; also usable in CI |
| Beta launch | August 5, 2026 |
| Out of beta | August 31, 2026 |
| Official platforms | macOS and Linux |
| Models | Muse Spark 1.2 and 1.3, Muse Wonder, Muse Wonder Full (subscribers) |
| Context window | 1 million tokens (Muse Spark 1.3) |
| Subscriptions | $5, $15, and $50 a month at launch; a $200 plan is listed in October coverage |
| Source code | Proprietary |
What Muse Code actually does
Muse Code is not a chat window that hands you snippets. It works inside your project folder from the terminal and carries out whole tasks. Meta describes it as an agent that plans changes, writes code, and validates the results in large repositories.
The features that set it apart:
- Persistent background agents. Helper agents stay active for the whole session instead of starting fresh for each subtask. Meta says this cuts repeated context-gathering.
- A local event log. Every model call, tool run, approval, and edit is recorded locally. Meta says a crashed session can resume from its last step, though independent testing of this is limited.
- Parallel sub-agents in git worktrees. A fan-out mode runs several write-capable agents in isolated git worktrees, so they don’t overwrite each other’s files.
- Built-in planning commands. /plan produces a plan you approve before work starts, /grill stress-tests that plan, and /goal keeps the agent working toward a stated objective.
- Added at general release (August 31): messaging between sessions, multi-agent workflows, a rewind control, and a TypeScript SDK in developer preview for building custom agents.
Muse Code pricing: every way to pay
There are three ways to pay, and they differ by more than price.
| Option | Price | What you should know |
| Everyday Usage plan | $5/month | About 10 to 50 requests every 5 hours, depending on task complexity |
| High Usage plan | $15/month | About 3x the Everyday allowance |
| Power Usage plan | $50/month | About 10x the Everyday allowance; includes Muse Wonder Full |
| Maximum Usage plan | $200/month | Listed in October 2026 coverage; includes Muse Wonder Full |
| Standard pay-as-you-go (Muse Spark) | $1.25 input / $0.15 cached input / $4.25 output per 1M tokens | Meta states it does not train on these prompts or outputs |
| Contributor tier (Muse Spark) | $0.10 input / $0.002 cached / $0.20 output per 1M tokens | Meta may use your prompts and outputs for training |
| Muse Wonder via API | $4 input / $0.20 cached / $20 output per 1M tokens | Meta’s frontier model, released October 12, 2026 |
Muse Spark 1.3, released September 2, 2026, kept the same price as 1.2. Reasoning tokens are billed as output, so long “thinking” runs cost more than the visible answer suggests.
What a coding session actually costs
Price tables are hard to picture, so here is an illustrative session: a mid-sized task that reads 2 million input tokens and writes 200,000 output tokens. These numbers are an example, not a measured average; your real usage will vary.
| Model and tier | No caching | With 1.5M of the input cached |
| Muse Spark, contributor tier | $0.24 | $0.09 |
| Muse Spark, standard tier | $3.35 | $1.70 |
| Muse Wonder, API | $12.00 | $6.30 |
Three takeaways:
- Caching matters a lot. Agents re-read the same files constantly. In this example, caching roughly halves the standard-tier bill.
- Model choice dwarfs everything else. The same session costs about 3.5x more on Muse Wonder than on Muse Spark standard. Use Wonder for hard tasks, not routine edits.
- The contributor discount is huge. It is about 14x cheaper than standard in this example. That gap is the price Meta is paying for your code.
If you would run more than one or two sessions like this a month on standard pricing, the $5 plan is likely the cheaper starting point. Plan allowances are counted in requests per 5-hour window, not tokens, so test your own workload before committing.
The contributor tier: cheap tokens, real trade-off
The contributor tier is the most talked-about part of Muse Code. In exchange for prices of about $0.10 and $0.20 per million tokens, Meta may use your prompts and completions to improve its models.
Before you choose it, check four things:
- Ownership and contracts. If you write code for clients or an employer, you may not have the right to share it for training. Check your agreements first.
- Rate limits. Reported limits are 60 requests per minute versus 3,000 on standard, and 2.1 million tokens per minute versus 4 million.
- Availability. The tier is limited to select countries and runs on a rolling 5-hour usage window.
- Defaults. Early users reported the contributor tier as the default after install. Confirm which tier you are on before opening a private repository.
A practical rule: use the contributor tier for open-source work, learning projects, and throwaway prototypes. Use standard or a subscription for anything proprietary.
Muse Code vs Claude Code: the price question
| Muse Code | Claude Code | |
| Maker | Meta | Anthropic |
| Interface | Terminal and CI | Terminal and other surfaces |
| Cheapest subscription | $5/month | Included with the $20/month Claude Pro plan |
| Higher tiers | $15, $50 and $200 a month | $100 (Max 5x) and $200 (Max 20x) a month |
| Data-sharing discount tier | Yes (contributor tier) | Not part of the plans above |
Price is Muse Code’s clearest advantage. On quality, treat vendor benchmarks with caution: in Meta’s own launch charts, Claude Opus 5 led all three coding benchmarks shown. The fair conclusion is that Muse Code competes mainly on cost, so test it on your own repository before switching.
How to install Muse Code
- Check your system. Use macOS or Linux. Windows users report it works through WSL2, but that is not officially supported.
- Run the installer. Meta’s one-line command is curl -fsSL https://dev.meta.ai/install.sh | bash. It pipes a remote script into your shell, so read the script first if you’re cautious.
- Sign in. Use your Meta account in the browser, or set an API key as META_API_KEY.
- Choose your tier. Pick a subscription, standard or contributor, before you work on private code.
- Start it in your project. Open your project folder in the terminal and type muse.
Who should use Muse Code?
- Good fit: price-sensitive solo developers, open-source maintainers, students, and teams already comfortable with terminal agents.
- Test first: teams using Claude Code or Codex who want a cheaper second agent for routine tasks.
- Probably not yet: Windows-only shops, and companies with strict IP rules that can’t confirm which tier every developer is on.
Conclusion
Muse Code AI is one of the most affordable coding agents from a major AI lab, as Meta’s AI chief Alexandr Wang has pitched it, and it improved quickly in its first three months. The catch is the tier choice: the lowest prices come with training rights over your code. Start with the $5 plan or the standard tier, measure your real usage, and move to the contributor tier only for code you’re happy to share.
Comparing AI coding tools? See our full AI coding assistant comparison to find the best price-to-performance fit for your workflow.
Frequently asked questions
Is Muse Code free?
The tool is a free download, but usage is paid: per token on the API, or through a monthly plan starting at $5.
What models does Muse Code use?
Muse Spark 1.2 and 1.3, plus Muse Wonder. Muse Wonder Full is available only to subscribers on the $50 and $200 plans.
Does Meta train on my code in Muse Code?
Only if you use the contributor tier. Meta states it does not train on prompts or outputs at the standard tier.
Does Muse Code work on Windows?
Not officially. Supported platforms are macOS and Linux; some users report success through WSL2.
Is Muse Code the same as the open-source MUSE code agent?
No. The open-source “MUSE” project on PyPI is an unrelated tool. Muse Code AI is Meta’s proprietary agent.


