Meta just entered the terminal coding agent race, and it brought a pricing weapon. Muse Code launched August 5 as a public beta CLI agent powered by Muse Spark 1.2. Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens. Flip to the Contributor tier and those numbers collapse to $0.10 in and $0.20 out — roughly 21x cheaper. The catch: your code becomes Meta’s training data. That trade-off is the real story here, and it splits cleanly along a single line: are you building personal projects or shipping production software?
What Meta Muse Code Actually Is #
Muse Code is a terminal agent — not a plugin, not an IDE extension. It reads your codebase, edits files, runs commands, and spawns subagents. It runs on macOS and Linux (no Windows yet), installs with a single shell command, and requires no Node.js or Python. The model powering it, Muse Spark 1.2, is Meta’s third coding-focused model release in four months, with a 1-million-token context window.
The architecture underneath is more interesting than the usual “AI in your terminal” pitch. Muse Code keeps specialized background agents alive for the entire session rather than spawning helpers fresh for each task. For large jobs, it fans out to subagents working in parallel, each in an isolated git worktree stored under .muse/worktrees/
— your working branch is never touched. Every action is appended to a local event log before it executes, so if the agent dies 20 hours into a complex task, it resumes exactly where it stopped. The OS sandbox is enforced natively via Seatbelt on macOS and bubblewrap on Linux — write access to your workspace, read-only everywhere else.
These are real engineering choices. The crash recovery and worktree isolation in particular are features the competition should be paying attention to.
Muse Spark 1.2 Benchmarks: Second Place, With a Caveat #
Meta published its own benchmarks at launch. On Terminal-Bench 2.1, Muse Code scores 82.9% — behind Claude Opus 5’s 86.7%. On DeepSWE v1.1, it scores 59.3% versus Opus 5’s 65.0%. Second place in this field is meaningful; this is not a toy.
The caveat matters: every number in the launch is vendor-run. No independent verification existed at publication time. Meta’s previous model came in 3.8 points below its published figure when the Terminal-Bench team ran their own evaluation. Treat the scores as directional, not definitive.
The Muse Code Pricing Decision You Have to Make #
The Contributor tier is the most interesting and most dangerous thing about Muse Code. At $0.10 input and $0.20 output per million tokens, you are looking at a fraction of what Claude Code or Codex costs in compute. For open-source work, side projects, or anything where your code is already public, this is a serious value proposition.
For anything else, think carefully. The Contributor tier gives Meta permission to train on your prompts and completions — that means your code, your debugging sessions, the business logic you spent months building. “Security, data privacy is paramount, and it’s not something we will trade off for a lower-cost option,” said Ken Ringdahl, CTO at Emburse. If you work with proprietary code, client projects, or anything under NDA, the Contributor discount is not yours to take. Meta has also not clarified what “training” fully covers — whether it includes evaluation runs, red-teaming logs, or internal abuse review. Unanswered questions at that scope are not a minor detail.
The Trust Problem #
The launch landed with a thud in several developer circles. Sign-in requires a Meta account — Facebook or Instagram OAuth — a friction point that felt tone-deaf for a product aimed at developers who have been opting out of Meta’s ecosystem for years. The Hacker News thread two days post-launch had only a few hundred comments. Developers who want Muse Spark 1.2 are routing it through third-party tools like OpenCode rather than Meta’s own CLI. The launch also coincided with news that a Meta AI model had been accidentally given internet access and breached a company website during a security test. None of this kills the product. But it reveals a gap between Meta’s product thinking and its intended audience.
The Bottom Line on Meta Muse Code #
Muse Code is a technically serious terminal agent at a competitive price point. The persistent subagent architecture, crash recovery, and native sandboxing are legitimately well-engineered. The benchmarks are encouraging with appropriate skepticism applied. If you work on open-source or personal projects and want the cheapest capable terminal agent available right now, the Contributor tier is worth evaluating. VentureBeat’s technical breakdown goes deep on the async agent architecture if you want the full picture.
For production or commercial work, stick with your current stack. The price difference does not justify the data exposure, and Meta hasn’t earned the benefit of the doubt on this one yet.