Meta Launches Muse Code, Its First AI Coding Agent, to Compete With Claude Code and Codex

Meta has released Muse Code, its first AI coding agent, entering direct competition with Anthropic's Claude Code and OpenAI's Codex. The tool is available now in beta and pairs with Meta's newest model, Muse Spark 1.2.

The launch was announced by Alexandr Wang, who leads Meta Superintelligence Labs and oversees the company's foundation model work. Wang joined Meta in June of last year as part of CEO Mark Zuckerberg's effort to reset the company's AI strategy.

What Muse Code Does

"Muse Code is a terminal coding agent, like many of the other coding agents on the market," Wang said. Developers install it with a single command and use it to handle full software engineering tasks — planning changes, writing code, and validating results — similar to how Claude Code and Codex operate.

Muse Code runs alongside Muse Spark 1.2, a coding-focused model that succeeds July's Muse Spark 1.1. Unlike the prior version, Muse Spark 1.2 was trained in tandem with Muse Code itself, which Wang claims improves overall coding performance compared to training the model in isolation.

Pricing

Meta is pricing API access to match Muse Spark 1.1: $1.25 per million input tokens and $4.25 per million output tokens on a pay-as-you-go basis. Wang also described a "contributor tier" priced at more than 10 times cheaper than the pay-as-you-go rate. In exchange, developers using that tier opt in to sharing usage data to help improve the model — a practice Wang said is standard across competing coding agents.

Meta is also beginning to field requests for zero-data retention arrangements, in which the company would not use developer data to train or improve its models. Wang called this "a big enterprise feature that is important for folks," a capability enterprise customers commonly require before adopting coding tools at scale.

Positioning Against Rivals

Wang was direct about Meta's competitive strategy: the company is differentiating Muse Code and the Muse Spark model family on price, not on claiming best-in-class capability. That contrasts with how Anthropic and OpenAI have marketed Claude Code and Codex, which lean heavily on benchmark performance and frontier capability claims.

No specific benchmark scores for Muse Spark 1.2 were disclosed in Meta's announcement, and the company has not published head-to-head comparisons against Claude Code or Codex on standard coding benchmarks like SWE-bench or HumanEval.

What This Means

Meta entering the coding agent market signals that AI coding assistance has become a battleground large enough to justify a dedicated product line, not just a model feature. By competing primarily on price — undercutting with a contributor tier over 10x cheaper than pay-as-you-go — Meta is betting that cost-sensitive developers and enterprises will trade marginal capability gaps for lower spend, rather than trying to out-benchmark Anthropic or OpenAI directly.

The zero-data-retention option also suggests Meta is targeting enterprise adoption specifically, an area where Anthropic and OpenAI already have established footholds. Whether Muse Code gains traction will depend less on Wang's claims about improved performance from joint training and more on independent developer testing once broader beta access rolls out. Meta has not disclosed a timeline for moving Muse Code out of beta.