 |
| Meta Muse Code |
Meta has joined the AI coding agent competition with Muse Code, a tool that works in a terminal and was released in beta on August 5 2026. It is designed to handle broader software engineering tasks rather than simply suggesting a few lines of code. If you have used Claude Code or OpenAIs Codex CLI before Muse Code is, in the group but Meta is trying to attract developers mostly because of the price. This guide explains what Muse Code actually does and how much it costs, how it functions inside
What Is Muse Code?
Muse Code is Metas coding agent that uses artificial intelligence. The people at Meta Superintelligence Labs built it. They have a person in charge of intelligence and his name is Alexandr Wang. You can use Muse Code on computers that run macOS and Linux. It works from the command line. You do not need to download a program or add it to your coding software to use it.
To get started with Muse Code you just need to type one command. Then you can ask it to do things like plan how to make a change to some code write the code test it and check if it works like it is supposed to.
Muse Code is powered by Meta's Muse Spark 1.2 model. They updated it to make it better at coding. The people, at Meta trained this version at the same time they were building Muse Code. They also made a tool to help the model work better when it has to do many tasks in a row to complete a piece of code
Features
Full engineering workflow. Muse Code is not limited to suggesting code snippets. It can plan a change write the code run validation and report back on whether the task succeeded, similar to how Claude Code and Codex operate.
Persistent background subagents. Muse Code can run agents on a task at the same time. According to Meta these subagents work inside environments commonly separate git worktrees so they do not interfere with each other while handling different parts of a job.
Local event log. The tool keeps a record of model calls tool use, approvals and edits on your machine. If a session crashes or fails partway through Muse Code can resume from where it left off of restarting the entire task.
Built-in planning commands. Muse Code includes commands for generating a plan for a task stress-testing that plan and continuing work toward a stated goal, which gives developers some control over how the agent approaches a job.
Approvals and sandboxing. The agent runs with an operating system sandbox and an approval step, by default, which is meant to limit what changes it can make without a person confirming them first.
Pricing: Standard vs. Contributor Tier
This is the part of the launch that has drawn the most attention, because Meta's pricing structure is unusual.
Muse Code is billed through the Muse Spark 1.2 model API, and Meta offers two tiers:
Standard tier
- $1.25 per million input tokens
- $4.25 per million output tokens
- $0.15 per million cached input tokens
- Rate limits up to 3,000 requests and 4 million tokens per minute per team
- Meta states that prompts and completions on the Standard tier are not used to train its models.
Contributor tier
- $0.10 per million input tokens
- $0.20 per million output tokens
- $0.002 per million cached input tokens
- Lower rate limits: 60 requests per minute and roughly 2.1 million tokens per minute
- Requires opting in to allow Meta to use prompts and completions for training purposes.
- The gap between the two tiers is large. Input tokens are roughly 12 times cheaper on the contributor tier, and output tokens are roughly 21 times cheaper. For a rough example, processing 100 million input tokens and 10 million output tokens would cost about $167.50 on the standard tier, compared with about $12 on the contributor tier. That is a meaningful saving, but it comes at the cost of handing over your code and prompts for Meta's training pipeline.
Both tiers require a Meta account with billing details set up before you can run the tool. It is also worth remembering that Muse Code is still in beta, so pricing, rate limits, and terms could change. Always check Meta's current developer documentation before relying on these numbers for budgeting.
How Muse Code Works
Installation of Muse Code is really simple. You just need to type one command in the terminal on your macOS or Linux computer. Now Muse Code does not work on Windows and it does not have a graphical app so you have to do everything in the command line.
When you want to start using Muse Code you have to choose a model route. This means you need to connect your Meta account to Muse Code and then pick a pricing route, for Muse Spark 1.2. You can choose the route or the contributor route. The route you choose will affect how much you pay and what happens to your data. Your data might be used for training if you pick one of the routes.
Now you can give Muse Code a task to do. You just need to tell it what you want it to do like fix a bug or build a feature in your code. Muse Code will plan out the steps it needs to take. Then work on the code. It can even work on parts of the task at the same time by using subagents.
Muse Code does not just stop after it writes the code. It will run tests to make sure the code works properly. This is what a human developer would do to verify their work. Muse Code will tell you if the changes it made actually worked.
If something interrupts Muse Code while it is working on a task it can pick up where it left off. This is because it keeps a record of what it has done far so it does not have to start all over again. This is really helpful because you do not have to worry about losing your work if something goes wrong.
How Does It Compare on Performance?
Meta has published benchmark comparisons showing Muse Spark against models from other AI labs, including on the widely used SWE-Bench Verified test, which measures how well a model can resolve real software issues. Independent reporting on these charts has noted that Meta's own comparisons still show Anthropic's Claude Opus 4.6 and Google's Gemini 3.1 Pro ahead of the original Muse Spark model on that benchmark, with Meta not highlighting specific updated scores for Muse Spark 1.2 in its announcement.
This matters for a simple reason: token price is not the same as cost per finished task. If a cheaper model needs more attempts, more back-and-forth, or more corrections to solve the same problem, the actual cost of getting a task done can end up closer to that of a more expensive but more accurate tool. Anyone comparing Muse Code with Claude Code or Codex on cost should think in terms of cost per completed task, not just cost per token.
Muse Code vs. Claude Code and Codex
All three tools are in the general category: AI agents that work through the terminal. These tools can plan, write and check code changes throughout a codebase. They are not just autocomplete tools.
The main differences now are:
Interface. Claude Code and Codex have options for integrating with different editors and workflows. Muse Code, when it first came out was for the terminal and had no special app.
Pricing model. Muse Code has a contributor tier that gives prices but asks for training data rights.
Track record. Claude Code and Codex have been available longer, giving them more time to build a track record with developers. Muse Code is new and still in beta so it is not yet clear how reliable it is for kinds of projects.
Since this is a changing area any direct comparison of performance including the one above is just a moment, in time. It is an idea to look at each companys latest information and independent tests before making a choice based only on what they say about their tools.
Who Should Use Muse Code
Muse Code could be a good fit if:
- You are comfortable working entirely from the terminal.
- You want the lowest possible token cost and are working on code that is not sensitive or proprietary.
- You want to experiment with an early-stage agent from a major AI lab.
It may not be the right choice if:
- You need a graphical interface or IDE integration.
- You work with client or proprietary code and cannot agree to Meta's training terms, since the contributor tier's savings are tied directly to giving up data rights.
- You need a tool with a longer, more proven track record on complex, long-running projects.
Limitations to Keep in Mind
- No native Windows support and no desktop app at launch.
- The contributor tier's low pricing depends on giving Meta rights to your prompts and completions, which is a real trade-off for anyone working with private code.
- The tool is in beta, so bugs, pricing changes, and rate limit adjustments are likely as it matures.
- Benchmark comparisons published by Meta should be checked against independent testing rather than taken at face value.
Final Thoughts
Muse Code is Metas tool that tries to beat Claude Code and OpenAIs Codex by focusing on low cost instead of the most powerful features. The contributor plan is very cheap. The cheap price comes with a data‑sharing rule that may not fit all developers, especially those who work with client or private code. If price is your priority and Meta's data terms fit your requirements, Muse Code may be worth considering. If you want a well‑tested tool that works on many platforms it may be better to pause and see how Muse Code grows or keep using a trusted tool for now.
This article is based on Meta's Muse Code launch information and independent sources available through August 2026. Prices, availability and features might change as Muse Code leaves beta so look at Metas guides before you decide to buy.
Source:
