Muse Code Review (2026): Pricing, Features & Honest Verdict

Muse Code is Meta’s terminal coding agent, powered by the new Muse Spark 1.2 model and built for long-horizon work across large repositories. The persistent background subagents are the real differentiator. Best for: engineers on big codebases who want a cheaper Claude Code rival. Price: From $1.25 per million input tokens (free plan: no, $20 signup credit). Rating: 8/10.

What is Muse Code?

Muse Code is a command-line coding agent Meta released on 5 August 2026, alongside a new model called Muse Spark 1.2 that was co-trained with it. You install the CLI, point it at a repository, and describe what you want changed. It plans the work, edits across multiple files, runs your tests and iterates until the change holds together. If that sounds like Claude Code or Codex CLI, it is because Meta is going directly at them.

The design choice that separates it is persistent, asynchronous background subagents. Most terminal agents work as one loop: think, act, observe, repeat, and every new sub-task starts by re-reading the same files. Muse Code spawns subagents that stay alive across the session, hold their own slice of context and keep executing while the main loop does something else. The claimed payoff is less redundant file reading, less steering from you, and the ability to keep going on a task for hours rather than minutes.

Meta’s own stress test is the number worth remembering: Muse Code iteratively optimised GPU kernels over more than 1,000 tool calls across up to 24 hours on NVIDIA Hopper hardware, producing competitive gains against baseline Triton implementations for KDA and MLA. Whatever you think of vendor benchmarks, a 24-hour agent run that ends with working optimised kernels is a different category of claim from “it wrote a to-do app”. It landed at rank 4 on Product Hunt on 6 August 2026 with 242 upvotes.

Some context on why Meta is here at all. Meta has no consumer subscription business in coding to protect and no developer platform revenue to cannibalise, which means it can price aggressively and does. It also has an enormous internal monorepo, which is exactly the environment that produces an agent tuned for long-horizon repository work rather than for greenfield demos. That origin shows in the product. Muse Code is not trying to help you build a landing page from a prompt. It is trying to survive a refactor.

What Are Muse Code’s Key Features?

Persistent async subagents

The subagents do not spin up and die per task. They stay resident for the session, carry their own accumulated context, and run in the background while the primary loop continues. In practice this attacks the failure mode everyone knows: an agent that re-greps the same directory for the eighth time because it forgot what it learned in step three. Whether it holds up on your codebase depends heavily on repository shape, but the mechanism is aimed at the right problem.

Long-horizon execution

Muse Code is explicitly tuned for tasks that run for hours, not for one-shot edits. Meta tested it to 1,000-plus tool calls in a single session. Most agents degrade badly long before that, either drifting off the original goal or filling their context with noise. If you have ever tried to get an agent to do a real migration across 200 files, you know that current tools want you to break it into 30 chunks and babysit each one. This is the first mainstream agent explicitly designed to not need that.

Multimodal inputs into the terminal

Meta’s launch demo feeds an mp4 fly-through video of a house into the terminal and gets back a working vacation rental marketing and booking site. Video into a CLI is a strange interface and also a useful one: hand it a screen recording of the bug instead of writing three paragraphs describing it. How reliable this is outside a curated demo is unknown, and we would not buy on the strength of it.

A model co-trained with the agent that runs it

Muse Spark 1.2 was trained alongside Muse Code rather than being a general model the agent was later bolted onto. Meta also describes a self-improvement loop where earlier model versions generate training data for later ones. Whether that produces a measurable edge is impossible to verify from outside, and Meta’s launch post shows benchmark charts for Terminal-Bench 2.1, DeepSWE 1.1 and an internal coding bench without publishing the underlying numbers on the page. That is a mild irritation. Charts without figures are marketing, and we would rather see the table.

Whole-repository planning and validation

The loop is plan, implement, validate, across the full repository rather than the files you named. The agent coordinates multiple subagents per task and checks its own results. This is table stakes for the category now, but Meta is claiming it works at a repository scale where the alternatives start to fall over, which is where Muse Spark 1.2’s co-training with the agent is supposed to matter.

How Much Does Muse Code Cost?

The CLI is free to install. You pay for model tokens through the Meta Model API: $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. New developers get $20 in free credits at signup, which is enough for real evaluation rather than a token demo. There is also a contributor tier at $0.10 per million input, $0.002 per million cached input and $0.20 per million output, in exchange for letting Meta train on your data.

That contributor tier is the aggressive part. It is roughly 12x cheaper on input and 21x cheaper on output than standard, which puts it far below anything comparable. Read the trade honestly: you are paying with your code. For open-source work or throwaway prototypes that may be fine. For a client’s proprietary repository it is not a decision to make casually, and it is not a decision to make on someone else’s behalf. Standard-tier pricing sits in the same neighbourhood as frontier competitors, so the headline saving only exists if you take the data deal.

There is a second cost worth planning for. A long-horizon agent burns tokens by design. A 1,000-tool-call session is not a $2 session, and the per-million rates that look cheap on a spreadsheet stop looking cheap when the agent runs for six hours. The cached input rate of $0.15 per million is the number that saves you, because a persistent subagent re-reading its own context is exactly what caching is for. Watch your first few long runs closely and set a budget before you let one go overnight.

Who is Muse Code Best For?

Use Muse Code if you work in a large repository and your bottleneck is agents that lose the plot halfway through a big change. The long-horizon design is the point, and nothing else on the market is marketing itself this specifically at multi-hour autonomous work. It is also compelling for anyone doing performance or systems work, given the kernel optimisation result.

The clearest use case is the one nobody enjoys: large mechanical migrations. Moving a codebase off a deprecated library, changing an API signature used in 300 places, converting a test suite from one framework to another. These jobs are too big for a single agent context, too repetitive for a person, and too irregular for a codemod. They are the reason “long-horizon” is a feature rather than a slogan, and they are where you should point Muse Code first.

Skip Muse Code if you are on Windows, because it currently ships for macOS and Linux only. Skip it if you need a stable production dependency today: it is beta, from a vendor with a track record of deprecating developer products, and it is one release old. And skip the contributor tier entirely if you handle code you do not own. The standard tier is priced normally and the discount is not worth a conversation with a client about where their source went.

Best Muse Code Alternatives

Claude Code is the benchmark in this category and the most mature terminal agent available, with the deepest tooling ecosystem around it. Included with Claude Pro at $20/mo or Max at $100/mo, or billed per token via API. Pick it if you want the safe answer.

OpenAI Codex CLI is the closest structural competitor, open source and included with ChatGPT Plus at $20/mo. Strong at focused single-task work, less explicitly built for multi-hour autonomous runs than Muse Code claims to be.

Gemini CLI is Google’s entry, open source with a generous free tier tied to a Google account. The cheapest way to run a capable terminal agent if budget is the constraint, and behind on long-horizon reliability.

Cursor’s agent mode deserves a mention as the fourth option, though it is an editor rather than a terminal tool. From $20/mo. If your team already lives in Cursor, adding a second agent workflow in the terminal is a real cost in habit and attention, and Muse Code has to be meaningfully better at long tasks to justify it. On the evidence so far it might be. On the evidence so far it is also beta.

Final Verdict: Is Muse Code Worth It?

Muse Code is a serious product, not a me-too launch. Persistent background subagents and explicit long-horizon tuning address the failure everyone actually hits with coding agents, and the 24-hour kernel optimisation run is the most concrete evidence anyone has published that a terminal agent can sustain a hard task for a full day. The $20 signup credit and free CLI mean testing it costs you an afternoon.

The caution is about Meta rather than the code. This is a beta from a company that has shipped and shelved developer tools before, and the pricing structure quietly pushes you toward handing over your repository in exchange for a 90% discount. Install it, run it against something you would be happy to open-source, and judge the long-horizon claim on your own repository. If it holds, it is the most interesting alternative to Claude Code that exists. Just pay the standard rate on anything that is not yours to give away.