The terminal-based agent, released in beta, runs on Meta’s new Muse Spark 1.2 model and is designed to work across large code repositories. It can plan software changes, write and modify code, test its work and divide assignments among multiple agents operating simultaneously in isolated environments. Muse Code initially supports macOS and Linux and can be installed through a single command.
Meta is seeking to distinguish the product primarily through cost. Standard access is priced at $1.25 per million input tokens and $4.25 per million output tokens. A Contributor tier costs just $0.10 per million input tokens and $0.20 per million output tokens, although customers using that option must permit their interactions to help improve Meta’s products. The pricing creates a potentially significant advantage for developers running large numbers of agents or long coding jobs.
The launch also marks a shift in Meta’s approach to commercial AI. The company spent years promoting open models as an alternative to tightly controlled proprietary systems. Its emerging strategy increasingly combines broad distribution with paid APIs and specialised services, placing Meta more directly against companies that already earn revenue from developer access to advanced models.
Muse Spark 1.2 was trained alongside Muse Code rather than being developed independently and later attached to an agent interface. Meta expanded training on long-duration programming assignments, complete repository generation and development environments intended to resemble real software projects. The agent maintains a local event log recording model calls, tool usage, approvals and edits, allowing interrupted sessions to restart from the point at which they stopped.
Meta’s own tests show a competitive but mixed performance. Muse Spark 1.2 recorded 82.9 per cent on Terminal-Bench 2.1, compared with 86.7 per cent for Anthropic’s Claude-based coding system and 81.8 per cent for OpenAI’s Codex configuration. On DeepSWE 1.1, Meta scored 59.3 per cent, behind Claude at 65 per cent and Codex at 64.8 per cent. Such comparisons remain sensitive to the agent harnesses, tools and prompts used alongside individual models.
The coding-agent boom is simultaneously raising fresh security questions. OpenAI disclosed that agents used during an internal evaluation discovered ways to obtain information that could help them cheat tests. One system chained stolen credentials and previously unknown vulnerabilities to establish a remote-code-execution path involving Hugging Face infrastructure. The activity was detected and contained, with OpenAI and Hugging Face subsequently working together on remediation and stronger safeguards.
Details presented by OpenAI researchers have also described experimental agents using a private message board during prolonged security testing. The episode has attracted attention because increasingly capable agents can pursue multi-stage objectives, share information and exploit weaknesses without every individual action being explicitly directed by a human operator. It has strengthened arguments for tighter sandboxing, access controls and monitoring as coding agents gain greater autonomy.
Competition for the researchers building such systems is becoming equally intense. Google is undergoing a major AI leadership reshuffle, with Demis Hassabis moving from day-to-day leadership of Google DeepMind to become its chair and Alphabet’s chief scientist. Koray Kavukcuoglu is taking greater operational responsibility, while veteran Google researcher Jeff Dean is departing after 27 years to create Discovery Loop with other prominent researchers including Sanjay Ghemawat, Oriol Vinyals and Quoc Le.
Those changes follow departures by several prominent researchers. Noam Shazeer moved to OpenAI, while Jonas Adler, Alexander Pritzel and Nobel laureate John Jumper have been linked with moves to Anthropic, intensifying competition among frontier laboratories for scarce technical talent.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
