Meta Just Entered the AI Coding War With a Surprisingly Cheap Agent
Most people assume the AI coding-agent market is a two-horse race between OpenAI Codex and Anthropic Claude Code. But on August 5, 2026, Meta released Muse Code — a terminal agent that undercuts both on price by a wide margin. The short version: it costs as little as $0.10 per million input tokens in its training-opt-in tier, a fraction of what Codex charges. Here I'll walk through what this new agent does, how its pricing compares to OpenAI Codex and Claude Code, and who should try it.
What Is Meta Muse Code?
It's Meta's first dedicated agentic coding tool, released in beta Wednesday 5 August 2026. It's a terminal agent for macOS and Linux, installable with a single command, and it plugs into large codebases the way that Claude Code or Codex does. There's no dedicated graphical app — you run it from your shell.
Muse Code is powered by Muse Spark 1.2, Meta's newest coding model, which the company says was co-trained with the agent itself. That co-training is a genuinely new detail: the agent's behaviors and the model's outputs were tuned as one unit rather than bolted together later.
Why It Handles Large Repos (Sub-Agent Worktrees)
The most interesting technical detail is how it scales to big codebases. When a task is large enough, CEO Mark Zuckerberg said, Muse Code "fans out to separate sub-agents working in parallel in isolated worktrees." Your working copy is never touched — every sub-agent edits a private copy, and the results are validated before anything is merged back.
Zuckerberg says in testing it built six features for a game simultaneously "with no collisions." By comparison, Codex and Claude Code mostly serialize edits or leave the developer to coordinate branches by hand. This isolated-worktree model is the single biggest differentiator, and anyone who has untangled merge conflicts will understand why it matters.
The Pricing That Changes the Comparison
Below is the freshest pricing data available, announced the same day as the launch:
- Standard tier: $1.25 per million input tokens and $4.25 per million output tokens
- Contributor tier (you consent to Meta training on your prompts): $0.10 per million input and $0.20 per million output — up to 20x cheaper
- Rate limits: the Contributor tier caps at 60 requests per minute and 2.1M tokens per minute; the Standard tier allows 3,000 requests per minute and 4M tokens
Alexandr Wang, who leads Meta Superintelligence Labs, told the Wall Street Journal: "For a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective."
That matters in real terms. A heavy user generating 20M output tokens in a month pays roughly $85 for output on the Standard tier — but only about $4 on the Contributor tier. That is not data ChatGPT or a fast search will hand you, because the pricing shipped the same week.
Muse Code vs Codex vs Claude Code: comparison
Here's the head-to-head:
| Feature | Muse Code (Contributor) | OpenAI Codex | Claude Code |
|---|---|---|---|
| Terminal agent | Yes | Yes | Yes |
| Output tokens cost (approx) | $0.20 / 1M | $10-$30+ / 1M depending on plan | ~$37.50 / 1M (Sonnet 4) |
| Parallel worktree sub-agents | Yes | No (serialized) | Limited |
| Training-opt-in discount | Up to 20x | No | No |
| Crash-safe event log | Yes | Partial | No |
Pricing as of early August 2026 per Meta's launch and MacRumors reporting; Codex and Claude rates are list prices and vary by plan.
Built-In Skills: /plan, /grill, /goal
Muse Code ships with default slash commands. /plan turns a task into an approval-gated plan you have to accept before code runs; /grill stress-tests an existing plan against edge cases; /goal keeps multiple sub-agents pointed at a single objective. The agent also keeps an append-only local event log of every model call, tool run, approval, and edit — if your computer crashes mid-task, it resumes from where it stopped instead of starting over.
Who Should Use It (and Who Should Wait)
If you are a solo developer or small team already paying for Codex or Claude Code, the Contributor tier at $0.10/$0.20 per million tokens is worth testing this week — the price alone is enough to run a real project through it. If your company cannot let Meta train on prompts due to compliance, the Standard tier is less remarkable, and Microsoft's / OpenAI's ecosystem integration and Anthropic's Claude Code are still strong choices.
For anyone needing enterprise approval flows or managed support, hold off — a curated arrangement has not shipped yet. Meta's been expanding its enterprise AI push (it entered the customer-service agent market in a previous quarter), so a managed option is plausible in the months ahead.
Frequently Asked Questions
Q: Is Meta Muse Code free to use?
A: Installing the beta is free, but usage is billed per token. The Contributor tier charges $0.10 per million input and $0.20 per million output tokens; the Standard tier charges $1.25 input and $4.25 output per million.
Q: Does Meta train on my code if I use Muse Code?
A: Only if you enable the Contributor tier, which lets Meta use your prompts and outputs to train future models. The Standard tier does not use your prompts for model training.
Q: What model does Muse Code run on?
A: It is powered by Muse Spark 1.2, Meta's purpose-built coding model, and runs out of the box. You'll only need to install the agent and provide an API key.
Q: Is Muse Code for developers only?
A: Mostly, yes — it is a terminal tool scripted for engineers. Non-technical users can trigger it with natural-language prompts, but you still need to review the code before shipping.
Q: How much cheaper is it than Codex or Claude Code?
A: In the Contributor tier, up to 20x cheaper than its own Standard tier: $0.10 vs $1.25 input and $0.20 vs $4.00 output — well below typical Codex and Claude estimates.
Summary
Meta Muse Code proves the coding-agent price war is far from over. Between dirt-cheap Contributor pricing, parallel worktree sub-agents, and a crash-safe event log, it's genuinely worth a try if you're burning money on Codex or Claude Code to cut the entire agent workflow. Sign up for the beta this week and put it through one of your real repositories.
Some links in this article may be affiliate links.