Meta has entered the AI coding assistant market with Muse Code, a new tool designed to compete directly with Anthropic's Claude Code and OpenAI's Codex by offering significantly lower pricing and persistent agent capabilities that can run uninterrupted for up to 24 hours. The beta release marks Meta's first dedicated coding agent and arrives as the company faces pressure to convert its massive AI infrastructure spending into revenue-generating products.

What Is Meta Muse Code and How Does It Work?

Muse Code is a command-line tool that lets developers hand off complete software engineering tasks without switching between separate applications. Unlike chat-based coding assistants that reset context with each request, Muse Code maintains persistent subagents that stay active throughout an entire session, allowing the system to coordinate multiple background processes simultaneously.

The tool runs on Muse Spark 1.2, a coding-focused model trained specifically alongside Muse Code itself. This co-training approach improved how well the model works with the agent's planning system, tool use, context management, and subagent coordination. Meta increased training compute dedicated to coding tasks and broadened the variety of environments the model trained on, including long-horizon jobs such as building entire repositories from scratch and completing large projects end to end.

What sets Muse Code apart is its ability to recover from crashes and interruptions. Every model request, tool execution, approval, and code edit gets written to a local event log that lets the runtime rebuild a session exactly as it was and resume work from the same point if the process crashes or gets interrupted.

How to Get Started With Meta Muse Code?

• Installation: Install Muse Code with a single command-line tool and begin taking on complete software engineering tasks across a wide variety of use cases.
• Planning Phase: Use the /plan command to turn a request into a structured plan that you must approve before work begins, ensuring alignment before the agent proceeds.
• Stress Testing: Run the /grill command to stress-test your plan for weaknesses and identify potential issues before the agent executes the full task.
• Goal Orientation: Use the /goal command to keep the agent oriented toward a specific objective across long sessions, maintaining focus on your intended outcome.
• Platform Access: Access Muse Code through macOS and Linux beta, with billing running through the Meta Model API and expected availability on the OpenRouter platform.

How Does Meta's Pricing Compare to Competitors?

Meta is competing primarily on cost rather than raw capability. Developers can access Muse Code through a pay-as-you-go plan priced at $1.25 per million input tokens and $4.25 per million output tokens, matching the API pricing of the earlier Muse Spark 1.1 release.

The more aggressive pricing comes through Meta's "contributor tier," which costs more than 10 times below the standard pay-as-you-go rate but requires developers to opt in and allow Meta to use their data to improve the model. For companies unwilling to share data, Meta is beginning to accept requests for zero-data retention, a feature described as important for enterprise customers.

"You can install it with one command and then use it to take on complete software engineering tasks across a wide variety of use cases, planning changes, writing code, validating the results," said Alexandr Wang, who leads Meta Superintelligence Labs and oversees the company's foundation model work.

Alexandr Wang, Head of Meta Superintelligence Labs

Wang joined Meta in mid-2025 as the centerpiece of CEO Mark Zuckerberg's effort to rebuild the company's AI strategy, and Muse Code represents his latest major product under that mandate.

What Can Muse Code Actually Do?

Meta demonstrated Muse Code's capabilities through a real-world test involving Nvidia hardware optimization. In one example, Muse Spark 1.2 used Muse Code to optimize GPU kernels for Nvidia Hopper graphics processors, running through more than 1,000 tool calls over sessions that stretched as long as 24 hours. The agent repeatedly wrote, compiled, profiled, and refined its own code, and Meta reported substantial performance gains compared with baseline implementations.

In another demonstration, a user fed the agent a video tour of a home, and Muse Code interpreted the footage to build a working vacation-rental marketing and booking website from it, showcasing the system's ability to translate unstructured input into functional code.

The persistent subagent architecture cuts down on repeated information-gathering and reduces latency during complex, multi-stage coding tasks. Rather than spinning up a fresh subagent for every request, Muse Code keeps subagents alive for the length of an entire session, allowing them to gather information, run additional steps, and decide on their own when to report findings back to the main agent.

Why Does This Matter for the AI Coding Market?

Meta entering the coding-agent space changes the competitive math for Anthropic and OpenAI, both of which have built meaningful developer businesses around Claude Code and Codex. By pricing Muse Code and Muse Spark aggressively, including a contributor tier Wang says is over 10 times cheaper than standard pay-as-you-go rates, Meta is signaling it intends to compete on cost first, not just on model performance.

The timing reflects Meta's broader financial pressures. The company generates roughly 98 percent of its revenue from advertising, and coding tools built on data-sharing incentives give it a new lever to offset the enormous capital spending going into AI infrastructure. The soft revenue forecast and shrinking free cash flow reported in Meta's most recent quarterly results added urgency to Zuckerberg's search for ways to turn heavy AI infrastructure spending into paying products.

Muse Code's beta is currently limited to macOS and Linux, with the Muse Spark 1.2 model also reachable through the Meta Model API with wider global availability. The model is additionally expected on the OpenRouter platform, which already hosts open-weight models from labs including DeepSeek and Z.ai, giving developers another route into Meta's coding stack.

" }