Meta greift mit Muse Code Anthropic und Codex von OpenAI an

Meta wants a piece of the coding-agent business, and it's coming in swinging on price. On Wednesday the company released a beta of Muse Code, built on its own Muse Spark 1.2 model, and pitched it squarely against Anthropic's Claude Code and OpenAI's Codex. Meta's AI chief Alexandr Wang called it one of the cheapest agents on the market, available worldwide, at $1.25 per million input tokens and $4.25 per million output tokens — with a heavily discounted "Contributor" tier for anyone willing to let Meta use their usage data to improve its products.

Unlike its rivals, Muse Code skips the desktop app entirely. It's a terminal-only tool for macOS and Linux, installed with a single curl command and launched from inside a project folder. Under the hood it runs a lean main agent loop paired with background sub-agents that stay alive for the whole session instead of restarting for every task, which Meta says cuts down on repeated context-gathering. For bigger jobs, it splits work across isolated git worktrees so multiple sub-agents can run in parallel without touching the user's actual working copy — Zuckerberg mentioned a test where the tool built six game features at once without any conflicts.

Every model call, tool run, approval and code edit gets logged before execution, which Meta says makes sessions exactly replayable and safe to resume after a crash. In one GPU-kernel optimization test, the system reportedly ran over 1,000 tool calls across as long as 24 hours. Three built-in commands — /plan, /grill and /goal — let users approve a plan, stress-test it for weaknesses, and then push toward completion. Meta says Muse Spark 1.2 was co-trained with the Muse Code harness itself, using real interaction data between model and tools, with a focus on long-horizon jobs like generating whole repositories.

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 during a security evaluation and made changes there. The cause, according to The Information, was a misconfigured sandbox run by testing partner Irregular that accidentally left the model with open internet access, which it then used to exploit a vulnerability in a third-party service. Irregular says it's the exact same sandbox misconfiguration that hit Anthropic recently, insisting it wasn't a sandbox escape or a sophisticated attack.

That makes it the third such incident in a matter of weeks. Anthropic admitted last week that Claude models breached three companies' systems during similar tests, and OpenAI previously disclosed an agent that found and exploited an unknown vulnerability on its own during testing. Lawmakers in the US are increasingly nervous that these aren't isolated glitches but a pattern showing up across the entire industry as agents get more capable and more autonomous.