Meta has stepped directly into the fight for AI-assisted software development, rolling out a beta version of a new tool built to go head-to-head with Anthropic’s Claude Code and OpenAI’s Codex. The Meta Muse Code agent is the company’s first dedicated coding assistant, and it arrives at a moment when Meta is under pressure to show that its massive AI spending can actually generate revenue.
Summary
Key takeaways
- Meta launched Muse Code in beta, a terminal-based coding agent built to compete with Claude Code and OpenAI Codex.
- The tool runs on Muse Spark 1.2, a coding-focused model trained alongside Muse Code itself.
- Muse Code can plan, write and validate code, while coordinating persistent subagents that stay active for an entire session.
- In one test, the system optimized Nvidia Hopper GPU kernels through more than 1,000 tool calls across sessions lasting up to 24 hours.
- Access starts with pay-as-you-go pricing matching Muse Spark 1.1 rates, alongside a cheaper contributor tier for developers who opt in to data sharing.
Meta launches Muse Code beta to challenge AI coding rivals
Muse Code is Meta’s answer to a market that Anthropic and OpenAI have largely defined so far. The agent lets developers install a single command-line tool and then hand off complete software engineering tasks, from planning changes to writing and checking code, without switching between separate applications.
The release comes from Alexandr Wang, who leads Meta Superintelligence Labs and oversees the company’s foundation model work. 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 is his latest major product under that mandate. “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,” Wang said in an interview reported by CNBC.
The timing matters. Meta’s shares dropped after the company issued a soft revenue forecast and reported shrinking free cash flow in its most recent quarterly results, adding urgency to Zuckerberg’s search for ways to turn heavy AI infrastructure spending into paying products.
Muse Code competes with Claude Code and OpenAI Codex
Functionally, Muse Code sits in the same category as Claude Code and Codex: it manages fleets of AI subagents that help build and maintain software inside one interface. Wang has framed the competitive strategy around price rather than raw capability, saying Meta is differentiating the Muse Spark model family “by price rather than capabilities” compared with what Anthropic and OpenAI offer.
Developers can access Muse Code through a pay-as-you-go plan that Wang says carries the same API pricing as the earlier Muse Spark 1.1 release: $1.25 per million input tokens and $4.25 per million output tokens. A cheaper “contributor tier” is also available, priced more than 10 times below the standard pay-as-you-go rate, but it requires developers to opt in and let Meta use their data to help improve the model. For companies unwilling to share data, Meta says it is beginning to accept requests for zero-data retention, a feature Wang described as important for enterprise customers.
Availability on macOS, Linux, and through the Meta Model API
Muse Code’s beta is currently limited to macOS and Linux. Meta says access and billing run through the same developer page that already hosts the Muse Spark API, and the Muse Spark 1.2 model itself is 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.
Core technology and capabilities behind the Meta Muse Code agent
What sets the Meta Muse Code agent apart from a simple chat-based coding assistant is its ability to run multiple background processes at once and remember exactly where it left off. Rather than spinning up a fresh subagent for every request, Muse Code keeps subagents alive for the length of an entire session.
These persistent subagents can gather information, run additional steps, and decide on their own when to report findings back to the main agent. Meta says this structure cuts down on repeated information-gathering and reduces latency during complex, multi-stage coding tasks — a meaningful difference from agents that reset context with every new instruction.
Local logging enables session recreation after crashes
Every model request, tool execution, approval, and code edit gets written to a local event log. That log lets the runtime rebuild a session exactly as it was and resume work from the same point if the process crashes or gets interrupted. For anyone running long, unattended coding jobs, that kind of continuity removes the risk of losing hours of progress to a dropped connection.
Built-in commands guide planning, testing, and goals
Muse Code ships with a handful of built-in skills designed to keep the agent on task. The /plan command turns a request into a structured plan that a developer has to approve before work begins. /grill then stress-tests that plan for weaknesses, and /goal keeps the agent oriented toward a specific objective across a long session. In one example Meta shared, 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.
Muse Spark 1.2: the model powering the new coding agent
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, and Meta says it brings real gains in code generation, complex debugging, and codebase understanding while keeping the broader agent skills intact. Meta trained the new model alongside Muse Code itself, an approach the company says improved how well the model works with the agent’s planning system, tool use, context management, and subagents.
To get there, Meta increased the training compute dedicated specifically to coding tasks and broadened the variety of environments the model trained on. Training runs included long-horizon jobs such as building entire repositories from scratch, completing large projects end to end, and running automated research tasks. The model relies on planning to organize its own work and on context compaction to hold onto important details across extended sessions. Meta also used the older Muse Spark 1.1 to generate difficult coding environments and instruction-following templates, then had it evaluate candidate solutions to produce extra training data for its successor.
A 24-hour GPU kernel test shows the agent’s staying power
Meta’s clearest demonstration of what long-session persistence can do involved Nvidia hardware. In one test, Muse Spark 1.2 used Muse Code to optimize GPU kernels for Nvidia Hopper GPUs, 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 the baseline implementations it started from.
Why this matters 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.
That strategy also fits into Meta’s larger financial picture. The company generates roughly 98% 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. Whether developers accustomed to Claude Code or Codex are willing to switch tools for lower prices, especially in exchange for opting into data collection, will likely shape how quickly Muse Code gains traction in a market where switching costs and existing workflows still matter.
FAQ
What is Muse Code?
Muse Code is an AI terminal coding agent launched by Meta in beta, designed to plan software changes, write code, validate results, and manage complex coding tasks.
What AI model powers Muse Code?
Muse Code is powered by the Muse Spark 1.2 AI model, which focuses on coding efficiency, debugging, and developer workflows.
How do Muse Code’s persistent subagents work?
Persistent subagents remain active throughout coding sessions to reduce latency and avoid repeated information gathering, coordinating complex tasks with less user input.
On which platforms is Muse Code available?
Muse Code beta is available on macOS and Linux, with Muse Spark 1.2 accessible through the Meta Model API for broader global use.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
