Meta Just Launched a Coding Agent. It's Cheap — But Read the Fine Print.

Every major AI lab now has a terminal coding agent. OpenAI has Codex. Anthropic has Claude Code. Google has Gemini CLI. And as of August 5, Meta has Muse Code — a direct competitor built on Muse Spark 1.2, a coding-focused model released alongside the agent.

The pitch is straightforward: install it with one shell command, point it at your repo, and let it plan, write, and validate code autonomously across large codebases. The pricing, though, is where things get interesting — and where developers need to read carefully before they start typing.

What Muse Code Actually Does

Muse Code is a terminal-based agentic coding tool — same category as Claude Code and Codex CLI. You run it in your terminal, give it a task, and it plans the changes, writes the code, and validates the results. It can handle full software engineering tasks across large repos, not just single-file edits.

The standout technical detail is how it handles parallel work. Muse Code spawns sub-agents into isolated git worktrees — similar to how Cursor 3 runs parallel agents — so it can work on multiple independent tasks simultaneously without file conflicts. It also ships with a crash-safe event log, which means if a session dies mid-task, it can resume from where it left off rather than starting over.

• Installs with a single shell command on macOS and Linux
• Runs parallel sub-agents in isolated git worktrees
• Crash-safe event log — resumes after interruptions automatically
• Sandboxed by default with approval prompts before executing commands
• Built on Muse Spark 1.2, a coding-specific model (not a general-purpose refresh)

How It Performs

Benchmark results put Muse Code in a clear middle position: ahead of Codex, behind Claude Code on Opus 5.

On Terminal-Bench 2.1, Muse Spark 1.2 scored 82.9% — ahead of GPT-5.6 Terra on Codex at 81.8%, but behind Claude Code on Opus 5 at 86.7%. On DeepSWE 1.1, Muse scored 59.3% versus 65.0% for Opus 5 and 64.8% for Codex. Even on Meta's own internal benchmark, Claude Code beat Muse by 8.8 points.

To be clear: these are good numbers. Muse Code is genuinely competitive. It's not the best terminal coding agent available, but it's in the conversation — and for many tasks, second-best at a fraction of the price is a perfectly reasonable trade.

The Pricing Model — and the Catch

This is where Muse Code gets genuinely interesting, and where developers need to pay attention.

Muse Code has two pricing tiers:

• Standard: $1.25 / $4.25 per million input/output tokens — competitive with the market
• Contributor: $0.10 / $0.20 per million input/output tokens — your prompts and code are used to train future Muse models

That's a 21x price difference. And here's the critical detail: Muse Code defaults to the Contributor tier after install. You have to actively switch to Standard to opt out of data collection — and the setting lives in pricing configuration, not in an obvious onboarding prompt.

For solo developers and open source projects, the Contributor tier might be a perfectly acceptable trade — $0.10 per million tokens for a capable coding agent is genuinely cheap, and contributing to model improvement is a reasonable exchange if you're not shipping proprietary code.

For enterprise teams, agencies, or anyone working with client code or sensitive IP: use Standard, set it explicitly in your configuration, and make sure anyone trialling the tool knows to do the same.

How It Compares to Claude Code and Codex

The honest comparison: Muse Code sits between Codex and Claude Code on raw capability, and undercuts both on price — even at Standard rates. The worktree-based parallel agent design is genuinely well-engineered, and the crash-safe resume is a practical detail that matters for long-running tasks.

Where Claude Code still leads: benchmark performance, ecosystem depth, and trust — particularly important in agentic contexts where the model is making decisions autonomously. Where Muse Code wins: price, and the $0.10 Contributor tier for non-sensitive work.

The interesting dynamic is that Meta releasing a capable, cheap competitor accelerates the whole category. OpenAI has already cut GPT-5.6 prices up to 80% in the same period. The terminal coding agent market is getting more competitive fast, and that's good for developers.

Worth Trying?

Yes — with the caveat. Muse Code is a genuinely solid agentic coding tool, and at Standard pricing it's competitive. The Contributor tier is an attractive option for personal or open-source projects if you're comfortable with the data trade. Just make sure you know which tier you're on before you start.

The terminal coding agent race now has five serious competitors: Claude Code, Codex, Cursor, Grok Build, and Muse Code. Trying a few of them on your actual workflow — not benchmarks — is the fastest way to figure out which one actually works for how you code.