If you read one line: Meta just shipped Muse Code, a terminal coding agent that installs with one curl line, runs on the new Muse Spark 1.2 model, and matches the Claude Code / Codex CLI playbook — with one twist worth pausing on: a "contributor" pricing tier that's roughly an order of magnitude cheaper in exchange for training on your code.
On August 5, 2026, Meta Superintelligence Labs released Muse Code — its first terminal coding agent — alongside Muse Spark 1.2, the model that powers it (Simon Willison, MarkTechPost). It's Meta's third model release in four months, and it drops straight into the category Claude Code and OpenAI's Codex CLI already defined. Here's what it is and whether it's worth your curl.
What Muse Code actually is#
Muse Code is a beta terminal coding agent for macOS and Linux (9to5Mac). It behaves the way you'd expect if you've used any of its rivals: it lives in your terminal, plans a change, writes and edits code across a repository, and validates the result. Two details set the shape of the product:
- Persistent async background agents. Muse Code can dispatch work to background agents that keep running while you do something else — Meta is leaning into long-horizon, coordinated tasks rather than one-shot edits (VentureBeat).
- A local event log. Every step the agent takes is recorded locally, so you can inspect and replay what it did. Built-in commands let you generate a plan, stress-test it, or tell it to keep working toward a goal.
It authenticates against the Meta Model API, which is OpenAI- and Anthropic-compatible — the same compatibility Muse Spark 1.1 shipped with — so if you already route through an SDK or a proxy, pointing at Muse is a base-URL change, not a rewrite.
How to start it#
Meta ships a one-line installer:
curl -fsSL https://dev.meta.ai/install.sh | bashStandard caution applies harder than usual here: this is a beta tool and a piped-shell installer. Read the script before you run it, and prefer a sandbox or a throwaway dev box over your main machine — the same hygiene we've argued for with every coding agent's shell access. Once installed, it points at the hosted Muse Spark 1.2 backend; there is no self-host option today (see the open-source note below).
The numbers: is Muse Spark 1.2 good enough to switch?#
Meta's headline is a coding-focused bump over 1.1:
- Artificial Analysis Intelligence Index: 54 — up 3 points from Muse Spark 1.1 (51) and 11 from the April 1.0 (43) (Artificial Analysis).
- Terminal-Bench 2.1: 82.9% and DeepSWE v1.1: 59.3% in Meta's own evaluations, up from 76.2% and 53.0% for 1.1.
That's a real, if incremental, gain — and it's competitive with the frontier coding agents on paper. But these are vendor-reported numbers on public benchmarks that models increasingly optimize toward. Treat them as a reason to try it, not a reason to switch. The only benchmark that decides your bill is your own repo; we walk through how to run that test in How to Evaluate an AI Coding Agent.
The pricing — and the catch#
Standard API pricing for Muse Spark 1.2 is unchanged from 1.1: $1.25 per 1M input tokens, $0.15 cached input, $4.25 output, with a ~1,048,576-token context. That already made 1.1 one of the cheapest frontier-class APIs, which we covered in Muse Spark 1.1 vs Kimi K3.
The new wrinkle is a "contributor" tier: roughly $0.10 input / $0.20 output per 1M — about an 8–10× discount — in exchange for permission to train future Meta models on your prompts and completions, capped at 60 requests per minute. That trade is the single most-discussed part of the launch, and it deserves its own decision, which we work through in When Meta's Contributor Discount Is Worth Your Code. Short version: it's a fine deal for throwaway and open-source work, and a bad one for anything proprietary or under NDA.
One caveat on all of these figures: they come from launch-week reporting and public commentary, not a Meta pricing page we could fetch directly (The Register). Confirm the exact rate on Meta's own docs before you hard-code it into a budget.
Open source: unresolved#
Meta built its reputation on open weights, so the obvious question is whether Muse Spark 1.2 or Muse Code will open. As of launch, neither had — and when asked directly, Mark Zuckerberg said only "I'll have more to share on that soon." Don't build a self-hosting or procurement plan on the assumption it will. Plan around the hosted API you can use today; if owning the backend is your actual requirement, Kimi K3's open weights remain the option that already ships.
The founder read#
Muse Code doesn't reinvent the terminal coding agent — it arrives as a credible third option in a category that's consolidating fast, with a genuinely cheap backend and a legitimately interesting async-agent design. If you're already running a three-agent stack, it's worth a slot on a non-sensitive repo this week. Two decisions gate it: run it on your code before trusting the benchmarks, and read the contributor tier as a data deal, not just a price — because that's exactly what it is.
