Meta Muse Spark 1.2 & Muse Code Release

Meta released Muse Code (beta), a terminal coding agent designed for long-horizon software engineering, alongside the Muse Spark 1.2 model. Both were jointly trained. Key details include: - Muse Code plans, implements, and validates complex multi-file changes across large repositories using persistent sub-agents that accumulate context and reduce latency. - It features parallel sub-agents working in isolated git worktrees to avoid conflicts, and an event log design enabling auditability, precise replay, and crash recovery. - Built-in skills include /plan, /grill, and /goal. - Installation is via a curl bash command followed by Meta account login. - Muse Spark 1.2 pricing: input $1.25 per million tokens, output $4.25 per million tokens, comparable to Gemini 3.6 Flash. - A data-contributor version (muse-spark-1.2-contributor) is priced at $0.10 input/$0.20 output, offering a roughly 12x discount for users who agree to data collection, matching DeepSeek-Flash pricing levels. - Access from China is heavily restricted, requiring specific network configurations.
Key insights
Meta Disrupts Pricing with Data-Collection Discount Model
Meta's Muse Spark 1.2 has a standard pricing tier ($1.25 input / $4.25 output per million tokens) but introduces a much cheaper contributor tier ($0.10 / $0.20) for users who allow Meta to collect their data for model training. This represents an approximate 12x discount. The strategy positions Meta as a price-disruptor, undercutting models like GPT-5.6 Luna and matching DeepSeek-Flash rates just as DeepSeek was rumored to raise prices, intensifying the AI pricing war.
Muse Code Features Persistent Agents and Parallel Worktrees
Muse Code is built for long-horizon software engineering tasks on large codebases. Its design includes persistent background agents that maintain context across sessions rather than restarting, reducing redundant exploration. For large tasks, the system distributes work to multiple sub-agents operating in isolated git worktrees to prevent conflicts, demonstrated by implementing six game features simultaneously without interference.
Crash Recovery and Auditability Through Event Logging
Muse Code incorporates an audit and recovery mechanism where every model call, tool execution, approval, and edit is written to a local event log before execution. This log becomes the single source of truth, allowing precise replay of sessions and exact recovery from crashes without losing progress or needing a new prompt.
Access Barriers for Chinese Users
Using Meta's models in China is described as extremely difficult. Users reported that standard virtual network cards do not suffice, and Meta actively blocks access, even preventing the website from loading, making usage more cumbersome than Anthropic's models in the region.