Meta introduced Muse Code in beta on August 5, 2026, marking the company's entry into the competitive field of AI coding agents. The tool, powered by the updated Muse Spark 1.2 model, is designed to handle complete software engineering tasks across large repositories, including planning changes, writing code, and validating results.

Mark Zuckerberg announced the release, describing Muse Code as a terminal coding agent that takes on complex work across large codebases. The agent coordinates multiple persistent subagents that operate in parallel within isolated worktrees, ensuring the main working copy remains untouched. In testing, the system simultaneously built multiple features for a game without collisions.

Key Features and Technical Design

Muse Code maintains a local event log that records every model call, tool run, approval, and edit. This creates a single source of truth, making the runtime replay-exact and restart-safe so the agent can resume precisely where it left off after a crash or interruption. Persistent background subagents remain active for an entire session, gathering information and executing steps independently before reporting back, which reduces repeated work and latency on multi-stage tasks.

Built-in commands guide the workflow. The /plan command converts a request into an approval-gated plan. /grill stress-tests the plan for weaknesses, and /goal keeps the agent focused on completing the stated objective over extended sessions. Meta co-trained Muse Spark 1.2 with the agent itself to improve coordination on planning, tool use, context management, and subagent handling.

The model represents a coding-focused update to Muse Spark 1.1. Meta increased training compute dedicated to coding tasks and expanded the variety of training environments. These included long-horizon projects such as building entire repositories from scratch and completing large end-to-end tasks. Older Muse Spark 1.1 was used to generate difficult coding problems and evaluate candidate solutions, producing additional training data.

In one demonstration, Muse Code optimized GPU kernels for NVIDIA Hopper chips. The process involved more than 1,000 tool calls over sessions lasting up to 24 hours, with the agent repeatedly writing, compiling, profiling, and refining code from a baseline without copying existing libraries. Another example showed the agent interpreting a fly-through video of a home and generating a functional vacation-rental website with booking capabilities.

Benchmark Results from Meta's Charts

Meta published comparisons showing Muse Spark 1.2 trailing Anthropic's Claude Opus 5 across the tested coding evaluations while outperforming OpenAI's Codex on several metrics.

Terminal-Bench 2.1, developed by the Laude Institute and Stanford researchers, involves 89 real-world tasks spanning system repair, data work, and security. Muse Spark 1.2 also edged Grok Build at 81.6% on that evaluation. DeepSWE 1.1 covers 113 coding tasks with internet access disabled during grading. Meta's internal benchmark draws from 440 real pull requests by its own engineers covering bug fixes, feature development, refactoring, and related work.

On a long-horizon GPU kernel optimization task involving over 1,000 tool calls, results placed Muse Spark 1.2 in the mid-pack for performance gains versus baseline. Independent leaderboard data for Terminal-Bench 2.1 has shown higher scores for other configurations of leading models, including figures above those Meta reported for Opus 5.

Meta measured against GPT-5.6 Terra for OpenAI comparisons in the primary coding charts rather than stronger variants. The company described Muse Spark 1.2 as a step toward the frontier, with larger and more capable models still in development.

Pricing Strategy and Availability

Alexandr Wang, who leads Meta Superintelligence Labs after joining through the Scale AI acquisition, has positioned the offering around cost rather than matching peak capabilities. Standard pay-as-you-go rates match the earlier Muse Spark 1.1 pricing at $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier is available at more than ten times lower cost for developers who opt in to allow Meta to train on their work. The company has begun accepting requests for zero-data retention options aimed at enterprise users.

Muse Code beta is available for installation on macOS and Linux via a single command. Muse Spark 1.2 is accessible through the Meta Model API with broader global availability and is expected on additional platforms.

The launch occurs as Meta continues heavy investment in AI infrastructure, with capital spending guidance reaching as high as $145 billion for the year. The company generates the large majority of its revenue from advertising and is seeking additional product lines tied to its AI development efforts.

Muse Code enters a market already served by Anthropic's Claude Code and OpenAI's Codex, along with other agentic tools. Its differentiators center on the crash-resilient event log, persistent subagents, and aggressive pricing rather than leading every published benchmark score.