Build8 publishers2 min readPublished
Meta hands its Muse models and coding agent to an enterprise unit run by CJ Desai
Meta grouped Muse API, Muse Code and Business Agent into a new enterprise platform whose Muse Spark model costs developers $1.25 per million input tokens. The endpoint is ready to test today, while the enterprise package still has no published price or delivery date.
The Engineer · Build desk
What happened
- Meta's developer overview says Muse models can be called directly or plugged into existing coding agents through OpenAI- and Anthropic-compatible interfaces.
- Meta's documentation describes Muse Code as a terminal and CI agent that plans tasks, edits files and runs commands, behind user approvals and an operating-system sandbox.
- Meta says Muse Spark 1.3 used roughly 20% fewer tool calls and 25% fewer tokens than its predecessor in comparisons run by its own engineers.
- A cheaper Contributor tier of the Model API lets Meta use customers' prompts and completions to train future models.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability A team with an agent harness built for OpenAI or Anthropic formats can, in principle, run Muse Spark against its own task set without rewriting the client, so a vendor bake-off is cheap to set up.
- decision Coding agents send the code they are working on as prompts, so taking the Contributor discount means deciding whether Meta may train on a team's source.
- constraint Until Meta prices and packages the platform, teams can benchmark the Muse Spark endpoint but cannot compare Meta's managed agent offer with rival vendors on cost or commitments.
Four product names are in play, and they are easy to tangle. Muse is the consumer agent, Muse Spark is the model, and Muse Code is the coding agent built on that model [4]. The first page of any evaluation doc will be a glossary.
The model endpoint is the piece a team can test now. Meta's Model API is generally available globally [6]. Muse Spark 1.3 is also offered through Oracle Cloud AI Platform, with Google Cloud access in private preview [16]. Oracle customers can run the trial inside an account they already hold [16].
The standard tier charges $4.25 per million output tokens against $1.25 for input [7]. Output costs 3.4 times as much as input [1]. Runtimewire noted that these are developer API prices, not a price for the broader Enterprise Platform [9].
Meta's efficiency figures need reading against that split. The token figure compares Muse Spark 1.3 with Meta's own previous model, measured by Meta's engineers [10]. For it to transfer, a team's tasks would have to look like the ones Meta ran. A saved output token is worth 3.4 saved input tokens at list price [1], and Meta reported one combined token figure [10].
Muse Code is where the engineering looks careful. User approvals plus an operating-system sandbox [11] is the right default for an agent that runs commands. Meta also says persistent background agents and a restart-safe event log let it handle work that runs longer than a quick coding prompt [12]. If the log works as described, an interrupted long job resumes from its record instead of starting from zero. The agent left beta at Connect and now runs on Windows [13]. The Decoder lists its competition as Claude Code, Codex and Cursor, plus cheap Chinese open-weight models [15].
The enterprise layer so far is a new unit with a leader. Zuckerberg described the effort as the "next major pillar" of Meta's business [1]. Desai said the platform will focus on "turning its AI stack into products and services that companies can deploy for their own businesses" [20]. Meta already reaches businesses at scale. It said in June that more than one million were using a Meta Business Agent on WhatsApp and Messenger [17]. Advertising supplied $196.18 billion of its $200.97 billion in 2025 revenue [18], or 97.6% [2]. The Wall Street Journal reported that Meta is spending more than $100 billion on AI infrastructure this year, according to The Decoder [19].
For a team whose agents already speak OpenAI or Anthropic formats [5], I'd run the endpoint trial now on the standard tier and revisit the platform once Meta prices it.
What to watch
- Whether Meta publishes enterprise pricing and packaging for the platform beyond the per-token Model API rates.
- Google Cloud availability for Muse Spark 1.3 moving out of private preview.
- Independent measurements of Muse Spark 1.3 token and tool-call use against rival models on agent workloads.