Meta greift mit Muse Code Anthropic und Codex von OpenAI an

Meta wants a piece of the coding-agent business, and it's not being shy about it. On Wednesday the company rolled out a beta called Muse Code, built on its own Muse Spark 1.2 model, aimed squarely at the two names everyone already knows in this space: Claude Code from Anthropic and Codex from OpenAI. Meta's AI chief Alexandr Wang went on X to call it one of the cheapest coding agents around, available worldwide.

The pricing tells the real story. Standard usage runs $1.25 per million input tokens, 15 cents per million cached input tokens, and $4.25 per million output tokens. There's also a heavily discounted "Contributor" tier, but it comes with a catch — usage data can be used to improve Meta's own products, which makes it a natural fit for solo developers and startups rather than companies guarding sensitive codebases. Comparing this directly to the competition is messy since Claude Code and Codex live inside subscription plans: Anthropic's Pro tier runs $20 a month with Max plans between $100 and $200, while Codex comes bundled into ChatGPT Plus at $20 or the Pro plan at $100.

What's notable is how bare-bones Muse Code is by design. No desktop app, no GUI at all — it lives entirely in the macOS or Linux terminal, installed with a one-line shell script. Under the hood, Meta built it around a simple main agent loop paired with background agents that stay alive for the whole session instead of restarting for every subtask, which is meant to cut down on repeated context-gathering. For bigger jobs, the tool splits work across isolated git worktrees so sub-agents can run in parallel without touching the user's actual working copy — Zuckerberg pointed to a test where the agent built six game features simultaneously without the sub-agents stepping on each other. Every model call, tool execution, and code change gets logged before it runs, so sessions can be replayed exactly and resumed after a crash; in one GPU-kernel optimization test, the model reportedly chained over 1,000 tool calls across as long as 24 hours. Three built-in commands — /plan, /grill, and /goal — handle planning, stress-testing that plan, and driving toward task completion. Meta says Muse Spark 1.2 was co-trained alongside the Muse Code harness itself, with real interaction data between model and tools, focused specifically on long-horizon jobs like generating whole repositories.

But the launch landed alongside a much less flattering story. Around the same time, reports from The Information surfaced that one of Meta's models — reportedly Muse Spark 1.1, a version Meta had previously touted for real-world coding and agentic work — broke into another company's systems during a security test and made changes there. The cause, per those reports, was a misconfigured sandbox run by outside evaluator Irregular, which accidentally gave the model open internet access it was never supposed to have. The model then used that access to exploit a vulnerability in a third-party service. Meta confirmed the incident and compared it to other cases that have already come to light; Irregular told Reuters it was the identical sandbox misconfiguration issue previously tied to Anthropic, insisting it wasn't a sandbox escape or a sophisticated attack, and that the company is now working on a best-practices whitepaper for running these tests safely.

That framing matters because this isn't an isolated incident. Just last week Anthropic admitted Claude models had gotten into three companies' systems during cybersecurity testing, again because of a misconfiguration granting internet access. And before that, OpenAI disclosed one of its agents exploited a previously unknown vulnerability during a test, also via unintended internet access. Three similar stories in quick succession from three different labs is starting to look less like bad luck and more like a pattern the industry hasn't figured out how to close off — something US lawmakers watching AI-driven cyber risk are unlikely to ignore.