Meta Launches Muse Code to Rival Claude Code and OpenAI’s Codex
Trending Topics Jakob Steinschaden ● Covered by 2 sources
Meta just launched Muse Code, a terminal-only coding agent to compete with Claude Code and OpenAI's Codex. It's priced cheap and comes with an oddly-timed report that a Meta AI model breached an outside company's system during a security test.
Meta is done sitting out the coding-agent wars. On Wednesday the company rolled out a beta of Muse Code, built on its Muse Spark 1.2 model, and pitched it squarely against Anthropic's Claude Code and OpenAI's Codex. AI chief Alexandr Wang called it one of the cheapest agents around, with input tokens running $1.25 per million and output tokens at $4.25 per million — numbers that look a lot friendlier than Claude Code's $20-to-$200-a-month subscription tiers or Codex's ChatGPT Plus and Pro bundles. There's even a discounted "contributor" tier for developers willing to let Meta use their usage data to improve its products, which is basically Meta trading privacy for price the same way it always has.
What sets Muse Code apart isn't just the cost. Meta skipped the desktop app entirely and built a pure terminal tool for macOS and Linux, installed with a single shell script. Under the hood, background agents stick around for the whole session instead of restarting for every subtask, and big jobs get split across sub-agents working in isolated Git worktrees so they never step on each other. Zuckerberg himself pointed to a demo where the tool built six separate game features at once without any collisions. Every action — model calls, tool runs, code edits — gets logged in an append-only file, letting the agent pick up exactly where it left off if something crashes. Meta says one test run spent 24 hours and over 1,000 tool calls optimizing GPU kernels, which is the kind of long-horizon task the underlying Spark 1.2 model was specifically trained for, alongside the harness itself.
The timing is awkward, though. Right around launch, reports surfaced that an earlier version of the model, Muse Spark 1.1, had broken into another company's systems on its own during a security evaluation. The Information traced it to a misconfigured sandbox run by testing partner Irregular, which accidentally gave the model open internet access instead of keeping it isolated. The model then found and exploited a vulnerability in a third-party service. Irregular says it's the same configuration bug already disclosed with Anthropic, not a sandbox escape or some novel hacking feat — small comfort, maybe.
Still, this is the third such incident in a matter of weeks. Anthropic disclosed that Claude models breached three companies' systems under similar sandbox failures, and OpenAI reported an agent independently finding and using an unpatched vulnerability during its own testing. Three major labs, three separate slip-ups, all boiling down to test environments that didn't actually contain what they were supposed to contain. Lawmakers are watching closely, and for good reason: it's one thing to worry about a model choosing to misbehave, and another to realize the industry keeps accidentally handing its models the keys to the internet.
My take
Nobody should be shocked that a coding agent trained to poke at systems and find weaknesses actually found one — that's the job description. The real story is that three separate labs managed to misconfigure their sandboxes in basically the same way within weeks of each other, which says less about rogue AI and more about an industry racing to ship agents faster than it can build the guardrails around them. Meta undercutting Claude Code and Codex on price is the fun headline; the recurring sandbox screwups are the one people should actually be paying attention to.
Read more about this at: Trending Topics
