Meta Superintelligence Labs Ships Muse Code to Take On Claude Code and Codex

Meta Superintelligence Labs finally ships a coding agent

Meta Superintelligence Labs shipped its first dedicated coding agent this week, and the timing says almost as much as the product itself. Muse Code, now available in beta for macOS and Linux, is a terminal-based ai coding agent built to handle complete software engineering tasks across large repositories: planning changes, writing code, and validating the results before handing them back to a developer. It installs with a single command.

The agent runs on Muse Spark 1.2, a coding-focused update to Meta's Muse Spark model family. Meta says the model and the agent were co-trained together, a design choice the company argues produces better tool use and fewer retries than bolting a generic model onto a third-party harness. Muse Spark 1.2 follows Muse Spark 1.1 from July and the original Muse Spark, which replaced Llama across Meta's consumer apps in April.

The headline feature is parallelism. When a job is big enough, Muse Code fans the work out to separate sub-agents that operate in isolated worktrees, leaving a developer's working copy untouched. Meta says internal testing had the agent build six features for a game at once with no collisions. An append-only local event log records every model call, tool run, approval, and edit, which Meta describes as replay-exact and restart-safe: if the agent crashes twenty hours into a long task, it resumes exactly where it stopped instead of starting over. Muse Code also ships three default skills: /plan turns a request into an approval-gated plan, /grill stress-tests that plan before work begins, and /goal drives the agent toward a stated objective.

Where Muse Code lands in a market Anthropic already owns

Muse Code enters a market that already has a clear leader. Anthropic's Claude Code has been estimated at roughly 54 percent of the AI coding market and grew from about 500 million dollars to an estimated 2.5 billion dollars in annualized revenue between September 2025 and February 2026, a ramp that outpaced nearly every enterprise software product before it. OpenAI's Codex and independents like Cursor make up most of the rest of the field developers currently choose from.

Against that backdrop, a natural search for any team evaluating options right now is a claude code alternative, and Meta clearly wants to be named as one. Alexandr Wang, the chief AI officer who leads Meta Superintelligence Labs, framed the pitch to the Wall Street Journal as a cost argument: Muse Code, he said, should be a strong option for many workflows largely because of price. Meta's standard API tier prices Muse Spark 1.2 at 1.25 dollars per million input tokens and 4.25 dollars per million output tokens, which sits close to rival mid-tier pricing rather than undercutting it outright.

The pricing tradeoff Meta is betting on

The real discount in Meta's setup shows up in a separate contributor tier: 0.10 dollars per million input tokens and 0.20 dollars per million output tokens, in exchange for letting Meta use those prompts to improve its models. Teams that want the lower price without giving up their code stay on the standard tier and pay closer to market rate instead. Muse Spark 1.2 is also available through OpenRouter for developers who want to swap it into an existing setup rather than adopt Muse Code itself.

There is no dedicated app interface yet, and no Windows build, gaps that Claude Code and Codex have already closed. What Muse Code has instead is a bet that pricing flexibility and a fully auditable run history will matter more to engineering teams than a polished front end, at least at this stage.

Why Meta needed a win here

Meta needed a win in this category. In June, the company pushed beyond its advertising-driven use of AI and entered the enterprise market with an agent built for customer service and daily operations, a signal that Meta wants AI to become a genuine second revenue line rather than a feature that keeps people scrolling. That ambition comes with real financial pressure: Meta narrowed its annual capital expenditure forecast in late July even as AI infrastructure spending kept climbing, after shares fell earlier in the quarter on a lighter revenue outlook and shrinking free cash flow.

A credible ai pair programmer with real developer adoption is one of the more direct ways to show that AI spending is turning into AI revenue, which is exactly the story investors have been pressing Meta to tell. Muse Code will not settle that question on day one, but it is the clearest evidence yet that Meta considers coding tools core infrastructure rather than a side project.

What US engineering teams should watch

• Which pricing tier you land on: the contributor tier is meaningfully cheaper but trades your prompts for the discount.
• No dedicated app or Windows build yet, so terminal-only teams get first access while everyone else waits.
• Whether isolated-worktree parallelism actually holds up on messy, real-world monorepos and not just Meta's own demos.
• How the replay-exact event log performs for compliance and audit needs versus Claude Code and Codex's own logging.
• Whether Meta backs this with the release pace Anthropic and OpenAI have kept up, rather than treating it as a one-off launch.

Sources