Meta Launches Muse Code, a Terminal AI Coding Agent Built to Take On Claude Code and Codex
Meta released the beta of Muse Code, a terminal-based AI coding agent powered by Muse Spark 1.2, undercutting Anthropic's pricing to compete with Claude Code and OpenAI's Codex.
Overview
Meta released a beta version of Muse Code this week, a terminal-based AI coding agent that the company is positioning against Anthropic’s Claude Code and OpenAI’s Codex, according to TechCrunch. The tool is Meta’s first coding-specific product built on Meta Superintelligence Labs’ newest model, and it arrives with aggressive pricing meant to undercut its rivals.
What We Know
Muse Code is currently available in beta for macOS and Linux and installs with a single terminal command, curl -fsSL https://dev.meta.ai/install.sh | bash, according to 9to5Mac. There is no dedicated graphical interface — the tool runs entirely through the terminal.
Meta CEO Mark Zuckerberg announced the release in a social media post, saying Muse Code can accomplish “complete software engineering tasks across large repos,” including “planning changes, writing code, validating the results,” according to TechCrunch. Muse Code was developed under Meta’s AI chief Alexandr Wang, who leads the Superintelligence Labs unit, according to SiliconANGLE.
The tool’s standout feature is how it handles large jobs. “When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees,” Zuckerberg explained. “Your working copy is never touched. In testing we had it build six features for a game simultaneously with no collisions,” according to TechCrunch. Muse Code also maintains a local event log that tracks model calls, tool usage, approvals, and edits, which is meant to allow a session to recover after a crash, according to 9to5Mac.
Muse Code is powered by Muse Spark 1.2, a coding-focused update to Meta’s proprietary model that provides “improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows,” according to Engadget. The original Muse Spark launched in April 2026 as Meta’s replacement for Llama, and Muse Spark 1.1 followed in July with added agentic and multimodal capabilities, according to 9to5Mac.
On pricing, Meta is offering Muse Spark 1.2 access at standard pay-as-you-go rates of $1.25 per million input tokens and $4.25 per million output tokens, according to Engadget, rates the outlet notes are “substantially lower” than Anthropic’s Sonnet 5 model, which costs $3 per million input tokens and $15 per million output tokens. Meta is also offering a cheaper “contributor” tier priced at $0.10 per million input tokens and $0.20 per million output tokens, which requires users to let Meta use their prompts and completions to improve its models, according to Engadget. Wang told CNBC that this contributor tier is “more than 10-times cheaper” than competing offerings, though it trades cost for data access, according to SiliconANGLE.
The launch continues Meta’s shift away from open-weight releases. As previously reported, Meta broke from its Llama strategy in April when it introduced the original Muse Spark as a proprietary, cloud-only model with no downloadable weights.
What We Don’t Know
Meta has not published independent benchmark comparisons showing how Muse Spark 1.2 performs against Claude Code or Codex on real-world coding tasks. It also remains to be seen how many developers will accept the data-sharing tradeoff required by the cheaper contributor tier, or whether enterprises will trust Meta with access to proprietary codebases given the company’s history with user data.
Analysis
Meta’s entry sharpens what has largely been a two-way contest between Anthropic and OpenAI in the terminal coding-agent category. By pairing a familiar feature set — parallel sub-agents, session recovery, single-command installation — with pricing well below Anthropic’s published rates, Meta is competing on cost rather than claiming outright technical superiority. The contributor tier’s data-access tradeoff also signals how Meta intends to keep closing the gap with rivals: subsidized access in exchange for training data, a strategy that mirrors the company’s broader pivot toward proprietary models since it retired its open-weight Llama approach earlier this year.