Meta has launched Muse Code, its first AI coding agent designed to compete directly with Anthropic’s Claude Code and OpenAI’s Codex.
The tool entered beta on Aug. 5 and is powered by Muse Spark 1.2, Meta’s latest AI model. It can write and debug code, run tests, verify its own work and handle longer software engineering tasks with less supervision.
Meta is also competing aggressively on price. Standard access to Muse Spark 1.2 costs $1.25 per million input tokens and $4.25 per million output tokens, giving developers another alternative in an increasingly crowded market for AI coding tools.
The bigger question is whether developers will trust Muse Code with the large, complicated projects where coding agents are supposed to provide the most value.
Muse Code Is Built For Longer Coding Tasks
Muse Code works from the terminal rather than operating only as a chatbot inside a browser.
That gives the agent access to the environment developers already use to navigate projects, edit files, run commands and test software.
Instead of simply generating a block of code and waiting for another prompt, Muse Code can work through multiple stages of a task.
It can inspect a codebase, make changes, check whether those changes work and continue refining the result.
Meta says the agent can also launch multiple sub-agents at the same time. That allows different parts of a larger problem to be handled in parallel rather than forcing one AI agent to complete everything sequentially.
For complex projects, that could make a noticeable difference.
A coding agent might have one sub-agent investigate a bug while another examines tests and another studies related parts of the codebase.
Muse Code Can Resume After A Crash
One of Muse Code’s more practical features is its persistent action log.
The system keeps track of the work it has already completed during a coding session.
If the process crashes or gets interrupted, Muse Code can use that record to continue from where it stopped rather than beginning the entire task again.
That sounds like a relatively small feature, but reliability becomes increasingly important as AI agents take on longer jobs.
Generating a function may take seconds. Refactoring a large application, investigating several failing tests or completing a software migration can involve dozens or hundreds of individual actions.
Losing all that context because of a failed session can remove much of the productivity advantage an autonomous coding agent is supposed to provide.
Meta Trained The Model And Agent Together
Muse Code is powered by Muse Spark 1.2, but Meta did not simply connect an existing model to a terminal interface.
The company says Muse Spark 1.2 and Muse Code were trained together.
That approach is intended to make the underlying model better understand how the coding agent operates, including how it uses tools, delegates tasks and verifies results.
It also continues the direction Meta established with Muse Spark 1.1.
That model emphasized agentic workflows, tool use, computer control and coding rather than focusing entirely on traditional chatbot conversations.
Muse Code turns those capabilities into a dedicated product for software developers.
Meta Is Competing On Price
Pricing could become one of Meta’s biggest advantages.
Muse Spark 1.2 costs $1.25 per million standard input tokens, $0.15 per million cached-input tokens and $4.25 per million output tokens.
That gives developers a usage-based alternative to coding products commonly bundled into monthly AI subscriptions.
Meta is also offering a cheaper contributor tier.
There is an important trade-off.
Activity from users on that tier may be used by Meta to improve its products. That could make the lower price appealing to individual developers working on personal or open projects, but businesses dealing with proprietary source code may need to look more closely at the data terms before using it.
For professional coding tools, privacy can matter almost as much as model performance.
A developer may be willing to share experimental code to receive lower prices. A company working on unreleased products, internal infrastructure or customer systems may not.
AI Coding Is Becoming A Major Battleground
Meta is entering a market where Anthropic and OpenAI already have considerable momentum.
Claude Code has become a major part of Anthropic’s push toward developers, while OpenAI has expanded Codex into increasingly autonomous software development workflows.
These tools are evolving beyond autocomplete.
Modern coding agents can search repositories, modify multiple files, run commands, execute tests, diagnose errors and continue working through a problem without requiring the developer to approve every individual action.
That makes software development one of the clearest commercial uses for increasingly capable AI agents.
Developers can directly measure whether a tool saves them time.
If an agent completes a debugging task in 10 minutes that previously took an engineer an hour, the value is easier to demonstrate than with many general-purpose AI features.
Muse Code Still Has To Prove Itself
Muse Code is currently in beta, which means benchmark results and feature lists will only tell part of the story.
The real test will be how reliably it performs inside large production codebases.
Coding agents need to do more than generate technically correct code. They need to understand existing architecture, avoid breaking unrelated features, follow project conventions and know when their proposed solution creates a new problem somewhere else.
Meta appears to be betting that tightly integrating Muse Spark 1.2 with Muse Code, adding persistent sessions and allowing multiple agents to work in parallel will make the system more useful for those longer tasks.
Price could help attract developers willing to test it.
But Claude Code and Codex already have established user bases, so Meta will need to prove that Muse Code offers more than simply another way to generate code from the terminal.
For now, its launch makes one thing clear: Meta wants a direct role in the AI coding race, and it is prepared to compete on both capability and cost.
via: CNBC | Reuters | Business Insider | Meta AI
