Meta Unleashes Muse Code: A New AI Coding Agent Powered by Muse Spark 1.2

Muse Code is Meta's answer to the demand for more autonomous and capable AI in software development. Unlike traditional AI coding assistants that primarily offer suggestions or answer programming questions, Muse Code is engineered to handle complete software engineering tasks. This includes everything from planning changes and writing code to validating results across large repositories. The agent operates directly within the terminal, making it a familiar environment for many developers, and is currently available in beta for macOS and Linux users.

A core differentiator of Muse Code is its architecture, which incorporates persistent asynchronous background agents. These specialized agents remain active throughout a coding session, building up context over time. This design choice aims to prevent redundant information gathering and reduce latency, particularly during multi-step tasks that can be cumbersome for other AI tools that restart their context with each interaction.

Another notable feature is its local event log runtime. Every action performed by Muse Code – including model calls, tool executions, user approvals, and code edits – is appended to a local, immutable log. This 'replay-exact and restart-safe' design ensures that if a session crashes or is interrupted, the agent can resume precisely where it left off without losing context or progress. This robustness is crucial for long-running, complex engineering projects where stability and continuity are paramount.

Muse Code also comes equipped with a set of bundled skills, accessible via simple terminal commands. These include /plan to convert a task into an approval-gated plan, /grill to stress-test a plan before implementation, and /goal to keep the agent focused on a defined objective. For larger projects, Muse Code leverages parallel sub-agents operating in isolated git worktrees. This allows multiple agents to work simultaneously on different parts of a project without conflicts, a capability demonstrated by Mark Zuckerberg himself when he used the tool to develop six game features concurrently without overwriting changes.

Powering Muse Code is Muse Spark 1.2, Meta's latest iteration of its proprietary Muse Spark family of frontier models. Released concurrently with Muse Code, Spark 1.2 is a coding-focused update to its predecessor, Muse Spark 1.1. Meta describes significant enhancements in code generation, complex debugging, and overall codebase understanding, specifically tailored for end-to-end developer workflows.

The model's improvements stem from substantially scaled-up training compute on coding tasks and expanded training environment diversity. Importantly, Muse Spark 1.2 was co-trained with Muse Code itself. This means the model was explicitly tuned to perform optimally within Muse Code's agentic harness, reflecting an industry trend away from treating models and tools as entirely separate entities. This co-training involved rejection-sampled harness trajectories and optimizations for goals, context compaction, and sub-agents, enhancing the model's ability to follow complex instructions and maintain long-horizon tasks.

Muse Spark 1.2 maintains a 1-million-token context window, a crucial feature for handling large codebases and extensive project contexts. While Meta has historically championed open-source AI through its Llama models, Muse Spark 1.2, like Muse Code, remains a proprietary offering, though Zuckerberg has hinted at future open-source possibilities for the Muse Spark family.

In terms of performance, Muse Spark 1.2 has shown competitive results on various benchmarks. On Terminal-Bench 2.1, it scored 82.9%, and on DeepSWE 1.1, it achieved 59.3%. While these scores place it strongly among current coding models, it often trails the absolute frontier models like Anthropic's Claude Opus 5 (e.g., 86.7% on Terminal-Bench 2.1 and 65.0% on DeepSWE 1.1) and OpenAI's GPT-5.6 Terra (64.8% on DeepSWE 1.1). However, Meta's internal coding benchmarks show Muse Spark 1.2 outperforming GPT-5.6 Terra and Gemini 3.6 Flash in some scenarios, highlighting its tailored optimization.

Perhaps one of the most intriguing aspects of Meta's Muse Code launch is its pricing strategy, which aims to be highly competitive and potentially disruptive. Muse Spark 1.2, accessible through the Meta Model API and powering Muse Code, offers a standard pay-as-you-go tier at $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million tokens.

However, Meta has introduced a heavily discounted 'contributor tier'. This tier offers significantly lower costs: $0.10 per million input tokens and $0.20 per million output tokens. The catch? Users on this tier explicitly grant Meta permission to use their prompts and completions to train future Meta models.

This strategy is a calculated move by Meta to acquire high-quality, real-world coding data, which is invaluable for improving AI models. Unlike publicly scraped data or synthetic data, user-generated code from active development environments provides crucial insights into real-world problems, reasoning processes, failures, and successful corrections. This approach echoes Meta's historical model of leveraging user-generated data for its social platforms, now applied to the developer ecosystem.

For individual developers and startups, this contributor tier presents an incredibly affordable entry point into advanced AI coding agents, potentially driving rapid adoption. However, enterprises dealing with highly proprietary or sensitive code may opt for the standard tier to maintain data privacy. This dual-tier approach allows Meta to cater to a broad spectrum of users while simultaneously fueling its AI research and development with a continuous stream of relevant data.

The launch of Muse Code and Muse Spark 1.2 underscores Meta's intensified commitment to AI monetization and its broader push into the AI agent paradigm. By offering a comprehensive, robust, and competitively priced coding agent, Meta aims to solidify its position in the rapidly evolving AI landscape and provide developers with powerful new tools to streamline their workflows and accelerate innovation.