Meta on the 5th (local time) unveiled its first artificial intelligence (AI) coding agent, Muse Code.
Meta Chief Executive Mark Zuckerberg directly announced the beta release of Muse Code on the social media platform X that day. He said it is a "terminal-based coding agent that handles the entire software engineering workflow, from planning changes across large code repositories to writing code and verifying results," and explained that a new model specialized for coding, MuseSpark 1.2, powers it.
Like Anthropic's Claude and OpenAI's Codex Assistant, Muse Code manages multiple AI agents and supports building apps within a single interface.
According to an internal performance comparison disclosed by Zuckerberg, MuseSpark 1.2 scored 82.9% on TerminalBench 2.1, which measures an AI coding agent's ability to perform real-world development tasks. That is ahead of OpenAI Codex (81.8%) but trails Anthropic Claude Code (86.7%). On DeepSWE 1.1, which gauges performance on unfamiliar tasks, it also lagged at 59.3%, behind Claude Code (65.0%) and Codex (64.8%).
In actual tests, MuseSpark 1.2 developed six game features simultaneously without causing code conflicts. In a "kernel optimization" task to speed up computations on graphics processing units (GPUs), it used Nvidia Hopper GPUs to try more than 1,000 times over 24 hours and identified improvements.
Meta highlighted a lower price than competing products as Muse Code's key differentiator. Zuckerberg said, "It's easy and inexpensive to install and start," adding, "With just a one-line install, you can start right away at the contributor tier."
Alexander Wang, who leads Meta Superintelligence Labs, said in an interview with CNBC that Muse Code follows the same pay-as-you-go pricing as the earlier MuseSpark 1.1. The price is $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.
Wang said the contributor tier is more than 10 times cheaper than that, and that Meta is differentiating itself from rivals with price competitiveness rather than performance.
Meanwhile, Meta also introduced a "zero data retention" option that does not use developer data to improve models. Wang described it as "a major feature important to corporate customers."
The release is seen as an attempt by Meta to expand revenue sources in its AI business while it pours massive funds into computing infrastructure such as data centers. Meta's share price plunged after a weak sales outlook and a decline in second-quarter free cash flow this year were disclosed last week.