Meta this week released Muse Code, a terminal-based coding agent designed to help developers organize and clean up their software projects. The tool runs on the company's new Muse Spark 1.2 model, which Meta says delivers stronger code generation capabilities than earlier versions. The beta release positions Meta as a direct competitor to OpenAI Codex and Anthropic's Claude Code in the race to automate software engineering work.
Muse Code functions as what Meta calls an "agent orchestrator" that operates from the command line. When a developer kicks off a task, the system launches background agents that keep a context file for other sub-agents to reference if they get off track. The platform logs every action before it runs, so no work gets lost. Multiple agents can tackle the same task in parallel using separate isolated work trees, and a developer's working copy stays untouched throughout the process. In one test, CEO Mark Zuckerberg said the platform built six features for a single game simultaneously with no conflicts between agents.
According to Zuckerberg's posts on X, Muse Spark took on a kernel optimization challenge running on Nvidia Hopper GPUs and made more than 1,000 tool calls across a 24-hour stretch. "It kept finding substantial improvements well beyond the initial exploration phase," he wrote. Hongyu Ren, a researcher at Meta's Superintelligence Labs, called it "a good harness" in his own X post. Benchmark tests using Terminal-Bench 2.1 and DeepSWE 1.1 showed Muse Spark performing near the top tier but not quite reaching first place when stacked against Opus 5, GPT5.6 Terra, Grok 4.5, and Gemini 3.6.
Meta's decision to keep Muse Spark as a closed-weight, cloud-hosted model marks a shift from the open-weight strategy it used with its Llama models. Zuckerberg didn't rule out opening the model later, telling a questioner on X, "I'll have more to share on that soon." Muse Spark 1.2 is the third release in four months from Meta Superintelligence Labs, a unit the company created in June 2025 to chase personalized AI "superintelligence" focused on deep reasoning and long-horizon planning. Meta researchers admitted with the first Spark release in April that coding abilities needed work, and version 1.2 fixes that gap. Developers can install the agent via a curl command or access Muse Spark through OpenRouter and its API. Pricing sits at $1.25 per million input tokens and $4.25 per million output tokens, with discounts available and a 1 million token context window. For companies evaluating AI-powered development tools, the choice between proprietary and open models now carries different implications depending on whether flexibility or cutting-edge performance takes priority. Meta's willingness to pivot its licensing strategy suggests the tradeoffs remain fluid even for major players.