Meta Muse Code: The AI Coding Agent Undercutting Claude Code and Codex

Published August 7, 2026 · Updated September 14, 2026By ABD Legacy LLC
AI tools

On August 5, 2026, Meta announced Muse Code, its first AI coding agent — a direct challenge to Anthropic's Claude Code and OpenAI's Codex. Meta's announcement was made by Mark Zuckerberg and detailed by Alexandr Wang, AI chief at Meta Superintelligence Labs. For agencies that run client work on coding agents, the headline is price: Meta is positioning Muse Code to compete on cost rather than capability.

What happened

The terminal-based agent runs on macOS and Linux, installs with one command, and is powered by the Muse Spark model family. The launch update shipped as Muse Spark 1.2; the lineup now runs to muse-spark-1.3, with muse-spark-1.2 and muse-spark-1.1 still available and contributor-tier variants of 1.3 and 1.2 alongside them — all reachable from Muse Code and the Meta Model API.

Meta's current product pages also describe Muse Code without a beta label — "a coding agent from Meta for the terminal and CI" — and the CLI has moved on since launch: version 1.1.1 adds OAuth 2.1 sign-in for MCP servers, todo reminders, and new built-in skills.

Muse Code targets end-to-end software engineering across large repositories: planning changes, writing code, and validating results. For larger jobs, it fans out to parallel sub-agents working in isolated worktrees, backed by a crash-safe event log and built-in skills such as /plan, /grill, and /goal.

What it costs

Muse Code is sold two ways. Pay-as-you-go remains at $1.25 per million input tokens and $4.25 per million output tokens, following the pricing introduced with Muse Spark 1.1 — per-token rates that undercut the incumbents at volume (see CNBC's launch coverage). Alongside it, Meta's Model API docs now document three flat monthly Muse Code subscription plans — Everyday Usage, High Usage, and Power Usage: a Muse Code subscription covers the CLI's auto-connected API key only, and other keys stay pay-as-you-go. The docs render the plan names but not the prices to unauthenticated readers, so treat the tiers as a structure decision and price your own token history against them (see the Muse Code subscriptions docs). A contributor tier offers lower rates in exchange for sharing training data, while enterprise customers can start with zero-data-retention.

To see how Muse Code pricing compares to the rest of the coding-agent landscape — flat per-seat plans, seat-plus-credits, and pay-per-token APIs — our AI coding agent pricing guide breaks down the billing models and cost variables to budget for.

Why agency teams should care

For agencies running client work on AI coding agents, pricing is the headline. Meta is positioning Muse Code to compete on price rather than capability, giving teams leverage in vendor negotiations and a cheaper option for high-volume coding work. The economics matter at scale: what cheaper coding agents do to agency margins is already reshaping how resellers and hourly shops price AI-assisted delivery.

The data trade-off matters too. The contributor tier exchanges training-data access for discounts, so agencies handling sensitive client code should review data-handling terms before opting in. The zero-data-retention enterprise tier is the compliance-friendly route.

How Muse Code compares to Codex and Claude Code

Muse Code is aimed squarely at Anthropic's Claude Code and OpenAI's Codex. Meta's stated differentiation is price: per-token rates undercut the incumbents at volume. On capability, Meta positions the agent as a comparable option for large-repo workflows rather than claiming a technical lead — the cost advantage is the pitch.

One Claude Code advantage to weigh: since v2.1.224 (August 2026), Claude Code sessions can message each other — short plain-text notes, automatic handoffs, and breakage warnings between parallel sessions (SendMessage / ListAgents). Muse Code's parallel sub-agents are orchestrated by the parent session, not peers messaging each other. If your agency runs parallel agent work, that difference matters. See what Claude Code's cross-session messaging means for AI agencies.

Next steps

Coding agents are commoditizing quickly, and a price war is a good moment to re-evaluate your toolchain. For practical comparisons of AI tools for agency work — coding agents, automation, and delivery tooling — browse findaiagency.com's AI tools coverage before committing your next project's stack.

Compare AI coding agents before you commit your next stack

Browse Vetted AI Agencies →

Or estimate your agency's coding-agent cost with the AI agency pricing calculator.

Frequently asked questions

What is Muse Code?

Muse Code is Meta's first AI coding agent. It is a terminal-based agent that runs on macOS and Linux, installs with one command, and runs on the Muse Spark model family, now at muse-spark-1.3. It targets end-to-end software engineering across large repositories and fans out to parallel sub-agents working in isolated worktrees for larger jobs.

How much does Muse Code cost?

Muse Code is pay-as-you-go at $1.25 per million input tokens and $4.25 per million output tokens, following the pricing introduced with Muse Spark 1.1. Meta's Model API docs also document three flat monthly subscription plans — Everyday Usage, High Usage, and Power Usage — covering the CLI's auto-connected API key only, with other keys staying pay-as-you-go. Plan prices are shown during onboarding rather than on the public docs page. A contributor tier offers lower rates in exchange for sharing training data, and enterprise customers can start with zero-data-retention.

Is Muse Code generally available?

Yes. Meta's current product pages describe Muse Code without a beta label, and the CLI is at version 1.1.1. It launched in beta on August 5, 2026 alongside Muse Spark 1.2, and Meta now documents flat monthly subscription plans alongside pay-as-you-go. At launch, Meta Superintelligence Labs AI chief Alexandr Wang declined to share user statistics.

Related reading

Sources