How is Muse Code different from Claude Code and OpenAI Codex?

Muse Code is a terminal-only tool, with no GUI, that Meta co-trained end to end with Muse Spark 1.2 so the model and the harness share the same instruction format. It bundles /plan, /grill, and /goal skills, runs async background sub-agents, and installs with a single curl command on macOS or Linux.

How much does Muse Spark 1.2 cost on the Meta Model API?

Muse Spark 1.2 stays at $1.25 per million input tokens and $4.25 per million output tokens on the pay-as-you-go tier — the same headline price as Muse Spark 1.1. A separate contributor tier is more than ten times cheaper for developers who agree to share their session data with Meta.

Is Muse Spark 1.2 faster on coding than Muse Spark 1.1?

Meta reports gains for Muse Spark 1.2 on Terminal-Bench 2.1, DeepSWE 1.1, GDPval, and its internal coding benchmark, and highlights a 24-hour GPU-kernel-optimization run of 1,000+ tool calls. Meta explicitly says it is not claiming benchmark superiority over Claude Code or OpenAI Codex.

Does Muse Code run on Windows?

Muse Code does not run on Windows at launch. The install script targets macOS and Linux, and there is no graphical application — the agent lives entirely in the terminal, which limits it to developers on those two operating systems for now.

What happens if the terminal crashes mid-task?

Muse Code writes a local event log of every step, so a crashed or restarted session can replay to the same state without losing plan, tool calls, or intermediate reasoning. Meta says the same log lets async background sub-agents run alongside the main agent inside a single session.

Meta Superintelligence Labs · 2026-08-05 · major

Muse Code + Muse Spark 1.2 — Meta ships a terminal coding agent

Muse Code is Meta's new terminal coding agent, powered by a co-trained Muse Spark 1.2 model. It runs on macOS and Linux, keeps a local event log so long tasks survive crashes, and ships /plan, /grill, /goal skills at $1.25/$4.25 per million tokens.

Meta's Muse Code + Muse Spark 1.2 announcement social preview

Meta's new Muse Code coding agent lives in the terminal and ships with a co-trained Muse Spark 1.2 model.

Key specs

Muse spark 1.2 input price	$1.25 / 1M tokens
Muse spark 1.2 output price	$4.25 / 1M tokens

Quick facts

Maker	Meta Superintelligence Labs
Products	Muse Code (terminal agent), Muse Spark 1.2 (model)
Install	curl -fsSL https://dev.meta.ai/install.sh \| bash
Platforms	macOS and Linux (no Windows, no GUI)
Availability	Public beta; Muse Spark 1.2 in Muse Code and Meta Model API
Bundled skills	/plan, /grill, /goal
Benchmarks reported	Terminal-Bench 2.1, DeepSWE 1.1, Meta Internal Coding Bench, GDPval

Pricing

Input (pay-as-you-go)	$1.25 / 1M tokens
Output (pay-as-you-go)	$4.25 / 1M tokens
Contributor tier · in exchange for sharing session data with Meta	~10x lower

source ↗

What is it?

Muse Code is a new terminal-based coding agent from Meta Superintelligence Labs. It plans changes, writes code, and validates results across large repositories from a shell prompt on macOS and Linux, with no GUI. Muse Code installs with one curl command from dev.meta.ai and comes bundled with three built-in skills: /plan, /grill, and /goal.

How does it work?

Muse Spark 1.2 was co-trained with Muse Code on long-horizon coding tasks, including whole-repository generation, so the model and the harness share one instruction format. The agent keeps a local event log for every tool call, which lets a crashed session replay to the exact same state and lets async background sub-agents run inside one session. Meta demonstrated the model optimizing GPU kernels over 1,000+ tool calls in a 24-hour run.

Why does it matter?

Muse Code puts Meta directly against Claude Code and OpenAI Codex on the terminal-agent turf where coding customers live. Muse Spark 1.2 stays at $1.25 input and $4.25 output per million tokens on the Meta Model API, well under Claude Opus 5, and a contributor tier drops that price more than tenfold for developers willing to share session data.

Who is it for?

backend and infra developers who live in the terminal

Frequently asked questions

How is Muse Code different from Claude Code and OpenAI Codex?: Muse Code is a terminal-only tool, with no GUI, that Meta co-trained end to end with Muse Spark 1.2 so the model and the harness share the same instruction format. It bundles /plan, /grill, and /goal skills, runs async background sub-agents, and installs with a single curl command on macOS or Linux.
How much does Muse Spark 1.2 cost on the Meta Model API?: Muse Spark 1.2 stays at $1.25 per million input tokens and $4.25 per million output tokens on the pay-as-you-go tier — the same headline price as Muse Spark 1.1. A separate contributor tier is more than ten times cheaper for developers who agree to share their session data with Meta.
Is Muse Spark 1.2 faster on coding than Muse Spark 1.1?: Meta reports gains for Muse Spark 1.2 on Terminal-Bench 2.1, DeepSWE 1.1, GDPval, and its internal coding benchmark, and highlights a 24-hour GPU-kernel-optimization run of 1,000+ tool calls. Meta explicitly says it is not claiming benchmark superiority over Claude Code or OpenAI Codex.
Does Muse Code run on Windows?: Muse Code does not run on Windows at launch. The install script targets macOS and Linux, and there is no graphical application — the agent lives entirely in the terminal, which limits it to developers on those two operating systems for now.
What happens if the terminal crashes mid-task?: Muse Code writes a local event log of every step, so a crashed or restarted session can replay to the same state without losing plan, tool calls, or intermediate reasoning. Meta says the same log lets async background sub-agents run alongside the main agent inside a single session.

Try it

curl -fsSL https://dev.meta.ai/install.sh | bash