Meta ships Muse Code, a Claude Code rival at $0.10/M contributor rate
macOS/Linux only, powered by the closed Muse Spark 1.2 with a 1M context — and that $0.10 rate means letting Meta train on your code.
Copy markdown
Meta walks into the terminal
Muse Code is Meta Superintelligence Labs' first terminal coding agent, out in beta on macOS and Linux — no Windows. A single command installs it (curl -fsSL https://dev.meta.ai/install.sh | bash), dropping it into the same lane as Claude Code, Codex, and Cursor.
The $0.10 tier bills you in data
Standard pricing is $1.25 in / $4.25 out per million tokens, but a 'contributor' tier cuts that to $0.10 / $0.20 if you let Meta train on your prompts and code. That undercuts DeepSeek V4 Flash ($0.14/$0.28) and GPT-5.6 Luna ($0.20/$1.20) — the price of the discount being your repo.
The model is welded to its harness
Muse Spark 1.2 is closed-weight, 1M-context, and reachable only through Muse Code and Meta's Model API. Meta co-trained it with the harness, and early users report tool-calling degrades when the weights run anywhere else — you're adopting the agent, not a model you can slot into your own stack.
Background agents, /plan, and crash recovery
The loop ships persistent background agents, a replay-safe event log that survives a mid-session crash, and bundled commands — /plan for approval-gated task planning and /goal to run to an objective. It's built for long-horizon work across large repos, not one-off single-file edits.
How Meta says it scores
Meta pits Spark 1.2 on Terminal-Bench 2.1 (89 tasks), DeepSWE v1.1 (113 tasks across 91 repos), and a 440-task internal PR bench against Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, Grok 4.5, and Kimi K3. These are vendor numbers, so treat them as a starting point — the real benchmark is your repo.