Meta launches Muse Code beta, undercutting Claude and OpenAI on price
Meta launched Muse Code on August 5, a terminal-based AI coding agent built on its new Muse Spark 1.2 model. Standard API pricing sits at $1.25 per million input tokens and $4.25 per million output tokens — well below Anthropic's Claude Sonnet 5, which charges $3 and $15 respectively. For developers already paying for AI-assisted coding, that gap is hard to ignore.
The agent
Muse Code runs entirely in the terminal. There is no dedicated app, no VS Code extension, no JetBrains plugin — a deliberate departure from how Claude Code and OpenAI Codex present themselves. Meta says the tool can write code from plain-text prompts, plan multi-step changes, spin up sub-agents to handle parallel tasks, and validate its own output. Muse Spark 1.2 was co-trained specifically with the Muse Code agent harness, meaning the model is tuned for long-horizon, repository-scale work rather than one-shot completions.
Demo material from Meta showed the agent building an interactive photon-sphere simulation, a Plants vs. Zombies-style game, and a webpage generated from an MP4 file.
The pricing tiers
Two tiers are on offer. The standard rate — $1.25/$4.25 per million tokens — applies to most users. A contributor tier drops that to just $0.10 input and $0.20 output, per CNBC, in exchange for sharing feedback that helps Meta improve the agent. Meta's chief AI officer Alexander Wang confirmed the tier structure directly. The contributor rate is more than 10× cheaper than standard Sonnet 5 output pricing — a significant pull for cost-conscious teams, though sharing proprietary code with Meta carries its own trade-offs worth considering, notes VentureBeat.
The caveats
The beta is available now for macOS and Linux via a curl install. Windows support has not been announced, which limits adoption in enterprise environments running mixed developer fleets. All published performance scores come from Meta's own evaluation harness — no independent benchmarks exist yet. And unlike Meta's Llama models, Muse Code runs on a closed-source model with no open-weights option, a notable shift for a company that built goodwill in the developer community on open releases. Whether that trade-off — lower prices in exchange for a closed, data-sharing model — makes sense depends on what you're building and who owns the code.
