Meta Launches Muse Code, a Terminal Coding Agent Powered by New Muse Spark 1.2 Model

Meta Enters the AI Coding Agent Race

Meta released Muse Code on August 5, 2026, its first dedicated AI coding agent, alongside a new model called Muse Spark 1.2 that powers it. Both are available now — Muse Code in beta for macOS and Linux, the model via the Meta Model API — putting Meta in direct competition with Anthropic's Claude Code and OpenAI's Codex for the fast-growing market of AI-assisted software engineering.

Mark Zuckerberg described it on social media as "a terminal coding agent that takes on complete software engineering tasks across large repos," positioning Muse Code not as a code completion tool but as a system that can plan, execute, and validate multi-step changes across an entire codebase.

What Muse Code Does

Unlike many AI coding tools that operate on individual files or functions, Muse Code is designed for repository-scale work. It ships with three built-in skills: /plan (generates a step-by-step implementation strategy), /grill (stress-tests that plan by probing for weaknesses and edge cases), and /goal (executes a specified target autonomously). The agent maintains an append-only local event log, which means it can crash and resume without losing context — a practical concern for long-running refactoring or optimization tasks.

Persistent async background agents are another notable feature. Rather than spawning a new agent for each subtask, Muse Code keeps background agents alive throughout a session, cutting redundant context-gathering and making multi-step workflows faster and more coherent.

Muse Spark 1.2: The Model Behind It

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. Meta says the model was co-trained with Muse Code itself, so the model's capabilities are specifically tuned to the agent's interface and task patterns — not adapted from a general-purpose model after the fact.

Benchmarks run on Terminal-Bench 2.1 (89 tasks), DeepSWE v1.1 (113 tasks across 91 repositories in 5 programming languages), and a Meta Internal Coding Bench (440 tasks from actual internal pull requests). Muse Spark 1.1 scored 80.0 on Terminal-Bench 2.1; Meta's documentation indicates 1.2 improves on that baseline, though exact scores were not disclosed at launch.

Pricing and Availability

Muse Code is available as a terminal application on macOS and Linux with a one-line install command. The underlying Muse Spark 1.2 model is available via the Meta Model API at $1.25 per million input tokens and $4.25 per million output tokens on a standard pay-as-you-go basis. Meta is also offering a heavily discounted contributor tier that cuts costs by more than 90% in exchange for data sharing — a pricing model designed to build developer adoption quickly while gathering training data.

Windows support has not been announced, and there is no dedicated graphical interface at launch — a gap compared to Claude Code and Codex, which both have richer integration options.

Why This Matters

The AI coding agent market is moving fast. Anthropic's Claude Code and OpenAI's Codex have strong footholds among professional developers, and GitHub Copilot remains the largest installed base. Meta's entry signals that the company sees agentic coding as a strategic layer it cannot cede to competitors, particularly as the lines between AI assistant and autonomous software engineer continue to blur.

Meta is not just making a product bet here. Muse Code is being built on top of the same Muse Spark model family that Meta is positioning as its frontier AI platform, which means the coding agent is both a product and a proving ground for more capable agentic systems to come. As reported by 9to5Mac and VentureBeat, the release marks Meta's most direct entry yet into the daily workflows of professional software developers.