Meta launches Muse Code, a coding agent aimed at professionals, offered via a monthly subscription ranging from $5 to $200. According to the test conducted, it matches Claude Code on bug hunting, but its visual output remains less impressive.
Since the initial successes of Llama, Meta had never returned to the forefront of generative AI. With the announcement of Muse Code in August, followed by Muse Wonder Full and Muse Wonder, presented as new frontier models (the most advanced at the moment) in October, the company now boasts, on paper, capabilities deemed credible for professionals.
Features close to the market, an attractive price
Muse Code offers the expected functions of a coding agent: goal, plan mode, sub-agents, and headless mode, which allows it to be controlled via script or by other agents. It also integrates MCP, hooks, skills, AGENTS.md, and computer use. It stands out with its sub-agents, launched in greater numbers by default and which remain active from one step to the next. However, there is no IDE extension, no desktop or mobile application, and no cloud agents.
The main advantage of the tool lies in its pricing. Monthly subscriptions come in four tiers: $5, $15, $50, and $200. Muse Wonder Full, described as “the most powerful model for code and reasoning,” is not included in the $5 and $15 plans. On the API side, billed per token (a unit of text read and produced by the model), the offering only includes Muse Wonder, at $4 per million input tokens, $0.20 for cache reads, and $20 for output.
Landing page and video: disappointing results
For the first test, a futuristic landing page for a fictional neo-cloud selling low-orbit datacenters, the agent often blocked non-sensitive operations, forcing manual validation of changes. The visual result is considered dated: the satellites spin too fast, continents are depicted by unconvincing blobs, and only the text remains coherent. On this same exercise, Google’s Gemini 4 “remains far ahead.”
The ten-second promotional video, created with Remotion, also failed to convince. According to the test, it resembles a slightly animated slideshow, with slides scrolling too fast to be read. Its aesthetic is judged as basic compared to productions from GPT-6, Claude Opus 5.5, or Gemini 4.
A more favorable bug hunt
The final test, inspired by the method of Paweł Huryn, a Polish product manager, involved fixing five bugs slipped into python-slugify, a Python library of about 200 lines. Muse Wonder Full and Claude Code with Claude Opus 5.5 were given the same copy, without Git history or internet access. Both tools fixed five bugs without breaking any tests, in 18.8 minutes for Muse Code versus 19.1 minutes for Claude Code. Muse Code consumed 5.54 million input tokens, compared to 4.96 million, but produced 65,000 output tokens, compared to 163,000 for Claude Code. It also spotted a flaw that the testers had not introduced.
According to the test, Muse Code does not yet have the versatility of Claude Code, Codex, or Antigravity, and lacks common sense as well as strong UI/UX capabilities. According to Artificial Analysis, Muse Code remains behind Gemini 4 with Antigravity in benchmarks (standardized performance tests). Due to the lack of per-token billing for Muse Wonder Full, the real cost of the tool remains impossible to estimate.
