Meta Finally Ships a Terminal Coding Agent
Meta released Muse Code on August 5: a terminal coding agent in public beta for macOS and Linux, one-command install, powered by the new Muse Spark 1.2 model (research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2). The Hacker News thread sat on the front page at 107 points. Meta is now officially in the Claude Code / Codex lane, a year after everyone else.
The interesting technical claim: Muse Spark 1.2 and Muse Code were trained together to operate as a pair. The model is a coding-focused update of Spark 1.1 with better debugging and codebase understanding, and the agent coordinates multiple persistent subagents per task across large repositories — planning, writing, validating. Co-training the model with its harness is the same bet Anthropic and OpenAI already made, and it is quietly becoming the industry consensus: the harness is not an afterthought, it is half the product.
Pricing is where Meta is actually competing. Pay-as-you-go is $1.25 per million input tokens and $4.25 per million output — a quarter of GPT-5.6 Sol's $5/$30. Then there is the contributor tier: more than 10x cheaper again, if you opt in to letting your usage improve the model. Read that twice. Meta is paying developers in compute for their coding data, the same trade that made Llama's ecosystem, now applied to agent trajectories. For a company that missed the coding-agent wave entirely, subsidized data collection at scale is a rational way to catch up — and for a lot of developers, 10x cheaper will be worth it.
← Back to all articles
The interesting technical claim: Muse Spark 1.2 and Muse Code were trained together to operate as a pair. The model is a coding-focused update of Spark 1.1 with better debugging and codebase understanding, and the agent coordinates multiple persistent subagents per task across large repositories — planning, writing, validating. Co-training the model with its harness is the same bet Anthropic and OpenAI already made, and it is quietly becoming the industry consensus: the harness is not an afterthought, it is half the product.
Pricing is where Meta is actually competing. Pay-as-you-go is $1.25 per million input tokens and $4.25 per million output — a quarter of GPT-5.6 Sol's $5/$30. Then there is the contributor tier: more than 10x cheaper again, if you opt in to letting your usage improve the model. Read that twice. Meta is paying developers in compute for their coding data, the same trade that made Llama's ecosystem, now applied to agent trajectories. For a company that missed the coding-agent wave entirely, subsidized data collection at scale is a rational way to catch up — and for a lot of developers, 10x cheaper will be worth it.
Comments