Meta Muse Code

Meta took Muse Code out of beta on August 31, 2026, less than a month after first shipping it, and the headline is the price. Three subscription tiers — Everyday Usage at $5/mo, High Usage at $15/mo, Power Usage at $50/mo — against Claude Code’s $20 Pro entry and Codex’s $20 Plus entry. That is a quarter of the going rate for the cheapest way in, and it is very obviously deliberate.

Muse Code is a terminal-based coding agent in the same shape as Claude Code and OpenAI Codex: you install it with one shell command, point it at a repository, and describe what you want changed. It runs on Meta’s Muse Spark 1.2 model. If you have never opened a terminal, stop here — this is not a visual builder and nothing about it is aimed at you. Start with Lovable or Replit instead.

What shipped at GA

Four things beyond the pricing:

Workflows. Orchestrate a team of specialised agents across a large job, carrying intermediate work between stages and returning one result at the end. There is a Workflows control room for watching and steering the agents while they run.

Inter-session messaging. Separate Muse Code sessions on the same machine can now share context and state when their work overlaps — useful if you habitually run three terminals at once.

Rewind. Roll the conversation and the code changes back to an earlier point in a session. This is a genuinely good feature and more agents should have it.

A TypeScript SDK, in developer preview, plus the underlying Muse Session Protocol. It lets your own software start and resume sessions, send instructions, and react to what the agent does, rather than requiring a human at a terminal. Anthropic and OpenAI both already ship equivalents, so this is catch-up rather than a lead.

The pricing is the product, and the pricing has a catch

Pay-as-you-go is still available at $1.25 per million input tokens and $4.25 per million output. There is also a Contributor tier at $0.10 in and $0.20 out — roughly a 12x and 21x discount — in exchange for opting in to having your prompts and completions used to improve Meta’s products.

Be clear about what that means. On Contributor pricing, the code you paste and the code the agent writes become training material. Several engineering leaders told The New Stack that this alone ruled it out for proprietary work, and they are right. If you are building something you intend to own, the Contributor tier is not a discount, it is a licence fee you are paying in the other direction. This is the same trade Lovable put in front of its Free and Pro users this month, and the same answer applies: read what you are agreeing to before the price talks you into it.

There is no free tier and no published trial credits. The $5 plan buys “10 to 50 requests every 5 hours,” which Meta admits varies with task complexity — vague, but no vaguer than Anthropic’s “5x Pro.”

Cheap is not the same as cheaper

The New Stack ran Muse Code and Claude Code through the same three coding tasks. Muse Code cost much less per run — and consumed substantially more tokens, produced a weaker refactor, and left dead code behind on another task. That is the number that matters and it is not on the pricing page. A $5 agent that needs three passes to do what a $20 agent does in one is not a saving, and you will not find out which you have until you are several hours in.

Verdict

Muse Code is a credible third entrant and the price war it started is good for everyone, including people who never install it. But Meta is behind, knows it, and is buying attention with margin — and Meta’s business is advertising, which makes the Contributor tier’s data trade worth reading twice rather than once.

Rating 2 of 5 for non-coders, and that is a ceiling, not a criticism: it is a CLI tool for people who already read diffs. If you are technical and already paying for Claude Code, the honest use case is a cheap second agent for bulk, low-stakes work — dependency bumps, test scaffolding, throwaway scripts on code you do not mind Meta seeing. Keep the work that matters on the tool that gets it right the first time.

Was this helpful?

YesCould be better