Meta Muse Code: A New AI Coding Agent Rivaling Claude Code and Codex

Meta shipped its first real coding agent this week — Muse Code, a terminal-based AI tool that plans, writes, and validates code across entire repositories, launched in beta on August 5, 2026 to go head-to-head with Anthropic’s Claude Code and OpenAI’s Codex. I’ve been testing agentic coding tools since the first wave of them showed up in terminals last year, and Meta’s timing here says a lot: the company waited until Claude Code and Codex had already trained developers to trust a CLI agent with their repos, then walked in with a price built to undercut both.

If you’ve used Claude Code or Codex, Muse Code will feel familiar in shape but different in one specific way — Meta built it to run several sub-agents on the same task at once, rather than one agent working a problem sequentially.

What Muse Code actually does

Muse Code installs with a single terminal command and runs on macOS and Linux. It’s powered by Muse Spark 1.2, a coding-specialized update to Meta’s Muse model family that was co-trained alongside the agent itself rather than bolted onto a general-purpose chat model after the fact. Mark Zuckerberg described it plainly on X the evening of August 5: it’s “a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results.”

The part that separates it from most agent CLIs I’ve tried is the architecture underneath. Muse Code runs persistent sub-agents in parallel, each isolated in its own Git worktree, so multiple lines of work can proceed on the same codebase without stepping on each other. A crash-safe event log tracks every action the agent takes, which matters more than it sounds — long-running agent sessions are exactly where you lose hours of work to a dropped connection or a bad `rm`. If a session dies mid-task, Muse Code can pick back up from the log instead of starting over.

The benchmark numbers, and why I’m not repeating them at face value

Meta published its own results on Terminal-Bench 2.1, DeepSWE 1.1, and GDPval, and early coverage framed those scores as putting Muse Code within striking distance of Gemini 3.6 Flash and xAI’s Grok Build 4.5. I want to be straight with you here: Meta did not release an independent, third-party-verified score alongside the launch, and outlets that dug into the numbers pointed out the obvious — a vendor’s own benchmark run on its own model is a marketing claim until someone outside the company reproduces it. I’d treat the DeepSWE and Terminal-Bench figures as “promising, unverified” until Artificial Analysis or a similar independent tracker runs its own pass. That’s not a knock on Meta specifically; every AI lab does this at launch, and it’s worth remembering every time a new model drops with a chart that conveniently beats the competition.

Pricing — this is the actual headline

Where Muse Code gets interesting isn’t the benchmarks, it’s the price. Meta is running two tiers:

The contributor tier is over 10x cheaper than standard, and the catch is exactly what you’d expect from Meta: opting in means your prompts and completions get used to train future versions of the model. For a solo developer or a small shop not touching sensitive code, that’s a genuinely appealing trade. For anyone working on proprietary or client codebases, standard pricing (or the zero-data-retention enterprise option Meta is also offering) is the only sane choice. Either way, $1.25/M input tokens on the standard tier undercuts what Anthropic and OpenAI charge for comparable coding-agent access, and that’s before you factor in the contributor discount at all.

Why Meta is doing this now

Coding agents have quietly become one of the most lucrative product categories in AI — not flashy consumer chatbots, but tools developers pay for every single day because they save real engineering hours. Anthropic’s Claude Code has been a meaningful chunk of the company’s enterprise revenue growth this year, and OpenAI’s Codex has similarly become a retention tool that keeps developers inside the ChatGPT ecosystem. Meta, which has spent 2026 trying to prove Llama’s successor models can compete commercially and not just on open-weight goodwill, needed a wedge into that market. Muse Code is that wedge, and undercutting on price while offering a genuinely different multi-agent architecture is a reasonable way to get developers to try it, even if the benchmark story is still unproven. It’s also a notable contrast with how fast Anthropic has been moving lately — see our coverage of the Claude Opus 5 launch — and it lands in a year where every major lab has had to reckon with what happens when these agents get too much autonomy, including OpenAI’s own sandbox-escape incident this summer.

It also plugs a gap in Meta’s AI story. The company has talked at length about Llama and Muse as research and consumer products, but it hasn’t had a serious answer to “which AI tool do your own engineers use to ship code” — a question Anthropic and OpenAI have been able to answer confidently for over a year now. Landing on OpenRouter and Meta’s own developer platform on day one suggests Meta wants distribution fast rather than building a walled garden first.

My take

I’ll be trying Muse Code on a real side project before I’d recommend switching a production workflow over from Claude Code or Codex — the worktree-isolated parallel agents are a smart design choice on paper, but “smart on paper” and “doesn’t silently corrupt your branch history” are two different claims, and only time and real usage settle the second one. What I will say is that the pricing alone makes this worth a look if you’re currently paying full freight for an agentic coding tool and don’t mind being an early adopter on the rough edges.

Frequently Asked Questions

What is Meta Muse Code?

Muse Code is Meta’s terminal-based AI coding agent, launched in beta on August 5, 2026. It plans, writes, and validates code across full repositories using parallel sub-agents, and is powered by the Muse Spark 1.2 model.

How much does Muse Code cost?

Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier drops that to $0.10/$0.20 per million tokens, but requires letting Meta use your prompts and code for model training. A zero-data-retention option exists for enterprise customers.

Is Muse Code better than Claude Code or OpenAI Codex?

It’s too early to say. Meta published its own benchmark scores on Terminal-Bench 2.1 and DeepSWE 1.1, but hasn’t released independently verified results, so any “better than” claim right now is coming from Meta’s own marketing rather than a third-party test.

What platforms does Muse Code support?

Muse Code installs via a single terminal command on macOS and Linux, and is available through Meta’s developer platform as well as OpenRouter.

What makes Muse Code technically different from other coding agents?

It runs multiple persistent sub-agents in parallel, each isolated in its own Git worktree, plus a crash-safe event log that lets a session resume after a dropped connection instead of losing progress.