Meta ships Muse Code, a coding agent with a full replay log
On 5 August, Meta's developer blog announced two things at once: Muse Spark 1.2, a model tuned for coding work, and an early beta of Muse Code, a coding agent that runs in the terminal. A coding agent is a program that reads your project, edits files and runs commands for you, step by step, instead of only answering in a chat.
Meta calls the new model a moderate improvement over Muse Spark 1.1. The more unusual part is the agent built around it.
An agent you can replay
Every subagent Muse Code starts, every tool call, and every time you steer or cancel it, goes into an event log that you can inspect and replay. When a big job splits into several tasks, Muse Code hands each one to a separate subagent, and each of those works in its own git worktree (a separate copy of the code on disk), so two agents never edit the same file at the same moment.
Meta says it trained the model together with Muse Code, and also across other agent tools, so it should still work in the coding agents people already use. The context window is 1 million tokens, which Meta says is enough to keep thousands of files of a project in one session. The model is available in Muse Code, through the Meta Model API and on OpenRouter.
Two tiers, and what each one costs
There are two ways to use it, chosen by model ID. The standard tier costs $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. The contributor tier is far cheaper on Meta's price table, at $0.10 per million input tokens and $0.20 per million output tokens, but it is limited by a token budget over a rolling five-hour window, is available only in some countries, and Meta says those prompts may be used to improve its products.
Muse Spark is a reasoning model, so the thinking it does before answering is billed as output. A setting called effort turns that thinking up or down. Meta also says it has started accepting requests for zero data retention, meaning your prompts are not kept after processing, through its sales team.
What to do before you try it
If you want to test it on code that belongs to your company or to a client, start on the standard tier, where Meta does not say your prompts are used for its products. Keep the contributor tier for personal or open source projects. The install is one terminal command, listed on Meta's post.
What we think
The replay log is the feature we would look at first. When an agent touches forty files, being able to go back and see which subagent changed what, and when, matters more than a few points on a benchmark. The pricing split is honest about the trade: a lower price in exchange for your prompts. Teams should decide that on purpose and not by default.
Sources