Meta enters agentic coding: Muse Spark 1.2 and the Muse Code terminal agent
In one sentence On August 5 Meta ships Muse Spark 1.2 (82.9% on Terminal-Bench 2.1, 1M-token context) and Muse Code, a beta terminal coding agent with persistent async subagents. The notable twist is a contributor tier: input at $0.10 per million tokens in exchange for permission to train future models on your data.
Meta has officially joined the command-line coding assistant race, the territory of Claude Code and Codex CLI. On August 5 Mark Zuckerberg announced Muse Code, an agent you launch from the terminal, hand natural-language instructions, and let write or edit code directly in your project, alongside the new version of the model powering it, Muse Spark 1.2.
The core idea matches the competitors: describe the task ("add tests to this module", "fix the cache bug") and the agent plans the changes, executes them, and validates the results. The difference Meta claims is parallel work management: Muse Code can coordinate multiple persistent subagents that keep working in the background on the same task, and every action lands in an observable, replayable log, handy when you need to reconstruct after the fact what the agent did and why.
The most debated move, though, is pricing. Beyond the standard tier, Meta introduces a contributor tier costing a fraction (up to 21x less on output): in exchange, you grant Meta permission to use your prompts and generated code to train future models. It is an explicit data-for-discount trade every team will have to weigh: a bargain for hobby projects, a near-automatic no for confidential corporate code.
For developers, Meta's entry mostly means more competition on price and features in a category that in 2026 has become the daily tool of millions.
Companies
Meta
Tools
Muse Spark 1.2, Muse Code
Tags
Sources