News Brief
What changed: Meta Muse Code, Meta's AI coding agent, completed a single engineering task in internal testing that ran for up to 24 hours and involved more than 1,000 tool calls. The agent left beta on September 1, 2026, and now runs on Meta's newer Muse Spark 1.3 model.
Why it matters: AI coding tools are moving from suggesting lines of code to working on whole projects for hours with little human input.
Who it affects: Software developers, engineering teams, and any business that pays for custom software, web apps, or automation.
Meta Muse Code was built to keep working long after a developer closes the laptop. In Meta's own testing, the AI coding agent stayed on a single graphics-processor optimization task for up to 24 hours, making more than 1,000 tool calls along the way. That one number has turned a routine product update into a bigger question across the software industry: if an AI can code for a full day without stopping, where does that leave the people who write code for a living?
Meta launched Muse Code in beta on August 5, 2026. Less than a month later, the company moved it out of beta, added multi-agent workflows and monthly plans, and switched it to a stronger model. Together, the moves signal that Meta is no longer experimenting with AI coding. It is competing for the developer market directly.
Inside the 24-hour test
The 24-hour figure comes from Meta's own internal testing, not from public users. Meta pointed Muse Code at a GPU optimization problem, the kind of slow, trial-and-error work that usually eats days of an engineer's time. The agent planned the job, ran experiments, checked results, and kept going for up to a full day, using more than 1,000 tool calls across the session.
What makes that possible is how Muse Code is built. A primary agent works toward the main goal while persistent background agents gather information and handle supporting tasks. Every model call, tool action, approval, and edit is written to a local event log, so if a long session crashes, it can pick up exactly where it stopped.
It is worth being clear about what the test does and does not prove. It shows the agent can stay on task for a long time. It does not show that the code it produced was better than an expert's, or that the same results will hold on every project. Meta has also required roughly 7,000 of its own engineers to use Muse Code every week, which means much of its early usage data is internal and mandated rather than organic.
How Muse Code got here
|
Date (2026) |
What happened |
|
Sept 2 |
Muse Spark 1.3 replaced 1.2 as the model behind Muse Code, with no price change |
|
Sept 1 |
Mark Zuckerberg announced Muse Code was out of beta, adding multi-agent workflows, sessions that message each other, an SDK preview, and monthly plans |
|
Aug 5 |
Meta launched Muse Code in beta for macOS and Linux, powered by Muse Spark 1.2 |
|
Apr 8 |
Meta Superintelligence Labs introduced the Muse Spark model family, replacing the Llama line |
The coding agent is also part of a larger Meta push under the Muse name. On September 8, Meta launched Muse, a consumer AI agent that climbed to the top of Apple's App Store, and on September 29 it added Muse for Small Business. Muse Code is the developer side of the same strategy.
What is Meta Muse Code?
Meta Muse Code is a terminal-based AI coding agent built by Meta Superintelligence Labs, the AI unit led by Meta's chief AI officer, Alexandr Wang. It is the lab's first product made specifically for coding. Developers install it with a single command and run it inside a project folder, where it plans changes, writes and edits code, runs commands, and checks its own results.
That makes it different from older AI coding assistants that mainly suggest the next line of code. Muse Code is designed to take a full engineering task, break it into steps, and carry it through to a tested result.
The technology behind it
Muse Code is the agent; Muse Spark is the model that powers it. Meta trained the Spark models together with the Muse Code harness, using long agent sessions as training material, so the model and the tool are tuned to work as a pair. Muse Spark 1.3 can hold about one million tokens of context, enough to keep a large codebase in view at once, and it accepts text, images, and video as input.
The features that matter most for long tasks:
-
Parallel sub-agents: Muse Code can split a big job across several agents, each working in its own isolated copy of the project, while the main working copy stays untouched. In a Meta demo, it built six game features at the same time this way.
-
Persistent background agents: Helper agents keep researching and handling side tasks while the main agent focuses on the goal.
-
Crash recovery: A local event log records every step, so long sessions can be resumed instead of restarted.
-
Safety controls: Approvals and an operating-system sandbox are switched on from the first run, so the agent asks before taking risky actions.
-
Automation-ready: A headless mode runs a single prompt to completion for scripts and CI pipelines, and a TypeScript SDK lets teams build Muse Code into their own tools.
Muse Code also reads project instruction files that teams already keep for other agents, including AGENTS.md and CLAUDE.md, which lowers the cost of switching from a competing tool.
Muse Code pricing: cheap, with a catch
Price is the center of Meta's pitch. Wang has openly described Muse Code as a strong option especially from a cost perspective, and the numbers back that up. Meta offers two token-based tiers through its Model API, and Spark 1.3 kept the same rates as 1.2.
|
Tier (per 1M tokens) |
Input |
Cached input |
Output |
Trains on your data |
|
Standard (muse-spark-1.3) |
$1.25 |
$0.15 |
$4.25 |
No |
|
Contributor (muse-spark-1.3-contributor) |
$0.10 |
$0.002 |
$0.20 |
Yes |
The Contributor tier is the catch. It costs a fraction of the Standard rate because Meta can use your prompts and completions to train its future models. For a hobby project, that may be a fair trade. For a business working with client code, private product logic, or regulated data, it is a decision that belongs with legal and security teams, not just developers.
Since leaving beta, Muse Code has also offered monthly subscription plans, which let developers skip per-token billing inside the agent.
Muse Code vs Claude Code vs Codex
Muse Code enters a market already led by Anthropic's Claude Code and OpenAI's Codex, and Meta is not claiming to beat them on raw capability. By Meta's own benchmark charts, Claude Opus 5 still leads on coding. Independent reviewers have reached a similar view: Muse Code competes well with the second tier of coding agents and trails the leader by a few points.
|
|
Meta Muse Code |
Claude Code |
OpenAI Codex |
|
Maker |
Meta Superintelligence Labs |
Anthropic |
OpenAI |
|
Interface |
Terminal and CI |
Terminal and IDE |
Terminal, IDE and cloud |
|
Main selling point |
Low price, parallel sub-agents, crash recovery |
Coding capability and maturity |
OpenAI ecosystem |
|
Track record |
New, out of beta since September 2026 |
Established |
Established |
Where Muse Code stands out is the combination of always-on background agents, an event log built for exact crash recovery, and aggressive pricing. Where it falls behind is maturity. Claude Code and Codex have longer track records and larger ecosystems around them, and for most teams the sensible move today is to test Muse Code, not to standardize on it.
So, should developers worry?
The short answer is no, but developers should pay attention. A 24-hour coding run does not mean AI can replace an engineer. It means the shape of an engineer's job is changing faster than many expected.
Agents like Muse Code are strongest at work that is long, repetitive, and easy to check: optimizing performance, migrating code, writing tests, fixing well-defined bugs. They are still weak at the parts that need human judgment: deciding what to build, understanding a client's real problem, weighing security and business risk, and owning the result when something breaks.
-
Defining the task clearly. An agent that runs for 24 hours on the wrong goal wastes 24 hours.
-
Reviewing and approving. Muse Code's approval system exists because someone still has to sign off on what the agent does.
-
Architecture and security. Faster code output means more code to review, and more places for mistakes to hide.
-
Data decisions. Choosing between the Standard and Contributor tiers is now part of a developer's job, because it decides who can learn from the company's code.
The developers most at risk are not the ones who use these tools. They are the ones who refuse to learn them while their competitors ship twice as fast.
What it means for businesses
For companies that buy software rather than build it, the impact will show up in timelines and budgets. Cheaper, longer-running coding agents can shorten the build phase ofweb applications, internal tools, and integrations. They also make it realistic for smaller teams to take on work that once needed a much larger engineering staff.
But faster code does not automatically mean better products. Businesses still need clear requirements, strong user experience, careful review, and a plan for where their data goes. The Contributor-tier tradeoff is a good example: a cheaper build that sends proprietary code to a third party for training may cost more in the long run than it saves.
There is also a search angle. As Meta, Google, and others put AI agents between customers and websites, how businesses are found online is changing too. Meta's consumer Muse agent already books, buys, and fills in forms for users, which means a company's site increasingly needs to be readable by AI agents as well as people. That makessearch engine optimization and structured, well-built websites more important, not less.
At American Design Hub, we see AI coding agents as a tool that speeds up delivery, not a replacement for strategy and craft. Teams exploringAI business automation or buildingtechnology products should test these agents now, set clear data rules, and keep experienced people in charge of the decisions that matter.
The bottom line
Meta Muse Code's 24-hour test is less a threat than a signal. AI coding agents can now stay on a problem for a full working day and longer, at a price that undercuts the market leaders. Claude Code still leads on capability, and Muse Code is still young. But the direction is clear: the developers and businesses that learn to direct these agents well will move faster than those that wait.
