Meta Muse Code: Terminal-Based AI Coding Agent With a $0.10 Contributor Tier

Meta entered the AI coding agent race on August 5, 2026, with the release of Muse Code, a terminal-based tool designed to compete directly with Anthropic's Claude Code and OpenAI's Codex . Powered by the new Muse Spark 1.2 model, the agent packs a 1-million-token context window and runs entirely in the developer's command line on macOS and Linux .

What Is Meta Muse Code?

Muse Code is an AI coding agent that runs inside your terminal. It was released in beta on August 5, 2026 and is designed to plan, write, test, and debug code across multiple background agents simultaneously . Developers install it with a single curl command:

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

The agent is the public-facing product from Meta Superintelligence Labs (MSL), the division led by Chief AI Officer Alexandr Wang, who joined Meta in June 2025 after Meta invested $14.3 billion in Scale AI .

The Muse Spark 1.2 Model

Muse Code is tightly coupled to Muse Spark 1.2, a coding-focused iteration of the Muse Spark model family . The Register reported that Meta co-trained Muse Code specifically on version 1.2, so the agent may not perform identically when run through other backends .

Key model specs:

• Context window: 1,048,576 tokens (1 million), retained from version 1.1 .
• Capabilities: multimodality, tool calling, structured output, and improved first-attempt coding accuracy .
• Persistent background agents: VentureBeat described it as tackling "24-hour coding jobs where other AI agents quit" .

The model is closed and proprietary, marking a shift from Meta's previous open-weight Llama series .

Pricing: The Contributor Tier Is the Headline

Meta is offering Muse Code through its Meta Model API in two pricing tiers . The standard tier is straightforward, but the contributor tier is where things get interesting:

Tier	Input (per 1M tokens)	Cached Input (per 1M)	Output (per 1M tokens)	Data Use by Meta
Standard	$1.25	$0.15	$4.25	No
Contributor	$0.10	$0.002	$0.20	Yes

The contributor tier is 12.5x cheaper on input and roughly 21x cheaper on output compared to the standard tier . However, in exchange for those prices, developers allow Meta to use their prompts and completions to train future models . The contributor tier is also rate-limited to 60 requests per minute (versus 3,000 per minute on the standard tier) .

Every new API account starts with $20 in free credits . According to CNBC, Wang described the pricing as "very aggressive and attractive" compared to comparable offerings from Anthropic and OpenAI . The standard tier alone is about one-quarter the cost of flagship rates from those rivals .

Competitive Positioning: Undercutting Claude Code and Codex

Muse Code enters a market dominated by Anthropic's Claude Code and OpenAI's Codex. By pricing the standard tier at roughly a quarter of its competitors' rates, and the contributor tier even lower, Meta is clearly aiming for volume adoption .

The launch follows a rapid cadence under Wang: Muse Spark 1.0 launched in early 2026, Muse Spark 1.1 arrived on July 9, 2026 (opening Meta's first paid API), and just 27 days later, Muse Spark 1.2 with Muse Code shipped on August 5 .

Internal AI Coding Targets at Meta

Meta is not just selling AI coding tools — it is also aggressively deploying them internally. According to internal documents reviewed by Business Insider and other outlets, Meta's "Creation Org" (covering Facebook, Messenger, and WhatsApp) set a first-half 2026 target requiring 65% of its engineers to write more than 75% of their committed code using AI tools . The Scalable Machine Learning team had a separate target of 50% to 80% AI-assisted code by February 2026 .

CEO Mark Zuckerberg has described this as part of making Meta an "AI-native company" . Reports indicate hundreds of employees were laid off in late March 2026 in connection with these AI-driven productivity goals .

Availability and Access

Muse Code is available exclusively through:

• Meta Model API (dev.meta.ai) — requires an API key.
• OpenRouter — a third-party AI model hosting platform.

There is no dedicated desktop or web app — the tool runs entirely in the terminal on macOS and Linux . U.S. developers got first access, but Meta has since expanded global availability of the Meta Model API .

Key Caveat: Co-Trained, Not Swappable

One important detail: Meta co-trained Muse Code specifically on version 1.2 of Muse Spark, meaning the agent is tightly coupled to that model release and may not perform identically when run through other model backends . The internal "MetaCode" agent referenced in some leaks appears to be separate internal tooling; Muse Code is the public-facing product built on similar principles .