The new AI agent runs in the terminal and offers autonomous tasks, with a data-sharing package priced at just $0.30 per million tokens, seriously undercutting rivals like Cursor and GitHub Copilot.
On August 6, 2026, Meta officially launched Muse Code, its first AI coding agent that works directly in the terminal. The tool is powered by Meta's latest model, Muse Spark 1.2, and is a beta release that takes on complete software engineering tasks-from planning changes and writing code to validating results.
"Releasing Muse Code in beta today. It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update," wrote Meta founder and CEO Mark Zuckerberg on X. In the announcement, Meta stated that the tool brings "improvements in code generation, complex debugging, codebase understanding, and overall end-to-end developer workflows."
Muse Code is not just another editor plugin. It runs on macOS and Linux as a command-line agent, meaning it fits into existing developer workflows without forcing a switch to a new editor. The key advantage is autonomy: the tool can run persistent background agents that continue working on tasks while the developer focuses on something else. This marks a shift from AI as an autocomplete tool to AI as a tireless junior colleague.
According to Meta's demonstrations, Muse Code can manage multiple sub-agents and delegate tasks, enabling the creation of functional software through text prompts. The company showcased examples like an interactive photon sphere model and a game inspired by "Plants vs. Zombies," as well as a video where the tool builds a website from an MP4 file.
However, Muse Code's sharpest weapon is its price. Meta offers a data-sharing package at just $0.30 per million tokens, a rate that seriously undercuts competitors for developers willing to let their code help train models. According to The Wall Street Journal, the contributor package will cost $0.10 per million input tokens and $0.20 per million output tokens.
For comparison, Muse Code's standard pricing is $1.25 per million input tokens and $4.25 per million output tokens. Competitor Anthropic's Sonnet 5 charges $3 per million input tokens and $15 per million output tokens. Alexander Wang, Meta's director of artificial intelligence, told CNBC that the contributor package offers a "significantly lower price" for users who agree to provide feedback.
Meta is entering a brutal but lucrative market. AI coding tools are the fastest-growing segment in the software industry, and the leading players are already massive. Startup Cursor, considered the main competitor, is reportedly raising funds at a $50 billion valuation after generating billions in revenue in just a few years. Alongside it are GitHub Copilot, OpenAI's coding tools, and Anthropic's Claude Code.
Competition is fierce, and everyone is grabbing for an edge. GitHub recently froze new sign-ups for Copilot because agentic usage broke through what their pricing could support. Meanwhile, Cursor is building its own rival to Anthropic's agentic coding product, showing that everyone is chasing the same vision of software that plans and executes tasks on its own.
Meta's strategy is clear: trade an extremely low price for access to real-world coding data. The data-sharing package is explicitly a trade for training data, meaning developers who save the most are also helping Meta catch up with more polished rivals. This bargain won't suit everyone, especially teams working on sensitive or proprietary code who don't want to share it with Meta.
The same logic has already proven successful for Chinese AI companies. Many firms have switched to models like DeepSeek to avoid the rising costs of using American AI tools. Meta is now applying a similar approach, leveraging its deep pockets and potential subsidization through data collection to offer a price that rivals will struggle to match.
There's also an awkward footnote. The same Muse Spark model family made headlines this week when an earlier version broke a security test and hacked an external company, a reminder that capable coding agents are also capable of mischief. Still, the launch signals Meta's serious intent to be relevant where the work actually happens-in the tools developers use every day.
