Meta launches Muse Code coding agent and upgraded Muse Spark 1.2 model

Meta AI Research on August 5, 2026 launched Muse Code, a beta terminal coding agent, alongside Muse Spark 1.2, an upgraded coding model that the company says improves on its predecessor across three benchmarks.

The Meta Muse Code launch pushes the company squarely into the developer-tools contest with Anthropic and OpenAI, and it leans on price to compete. Muse Code is a multi-agent harness: instead of one agent handling a job, it distributes work to asynchronous background sub-agents, logs events locally so a crashed session can recover, and ships with bundled skills for planning and stress-testing code.

Pricing is where Meta draws the sharpest line. The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens. A “contributor” tier cuts that to $0.10 and $0.20 in exchange for letting Meta train on users’ prompts and completions, an explicit data-for-discount trade.

On Muse Spark 1.2, Meta reported gains over Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash and Kimi K3 on Terminal-Bench 2.1 and its own internal benchmark of 440 tasks drawn from real pull requests. Those comparisons are Meta’s own and are not independently verified. The company also cautioned that its harness “may not be tuned for third-party models,” a caveat that cuts against reading the head-to-head numbers as neutral.

Muse Spark 1.2 is available now inside Muse Code, through the Meta Model API with expanded global access, and via OpenRouter. Whether the contributor tier’s price cut is enough to pull developers away from entrenched assistants, or whether the data-sharing terms scare enterprise buyers off, is the open question the launch leaves.