Verdict: Meta's Muse Code is a solid first entry into the AI coding harness race — it has voice input, sub-agents, skills, and a /goal loop that match Claude Code's feature set, plus session messaging that Claude Code does not offer. But the real story is pricing: the "contributor" tier of Muse Spark 1.2 costs $0.10 per million input tokens and $0.002 per million cached input tokens — roughly 12x cheaper than the non-contributor tier and dramatically cheaper than Claude or GPT-5.6. The catch: contributor mode means Meta can train on your data. For personal projects, experiments, and non-sensitive work, it is borderline free. For company code, you need the non-contributor tier, and the discount disappears.
Last verified: 2026-08-06 · Best for: Developers who want a cheap, capable coding agent and are OK sharing non-sensitive code with Meta · Volatile facts: Pricing, features, and model quality may change — this is a beta product.
- Muse Code is a terminal-based AI coding harness (like Claude Code or Codex CLI), currently in beta and not open source.
- Muse Spark 1.2 is the coding-optimized model that powers it, with a 1M-token context window.
- Contributor pricing ($0.10/M input, $0.002/M cached, $0.20/M output) is ~12x cheaper than non-contributor ($1.25/$0.15/$4.25) — but Meta trains on your data.
- Voice input is built in and works over API — a feature Claude Code only offers with a subscription, not on API.
- Features include: sub-agents, skills,
/goalloop, effort levels (ultra to minimal), web search (client/hosted/off), session messaging (experimental), themes. - The model is not frontier-tier. It built a working platformer game in one session but the presenter noted it is not as powerful as Claude Fable or GPT-5.6 Sol.
Pricing and model versions change often — last verified August 6, 2026.
What Is Muse Code?
Muse Code is Meta's terminal-based AI coding harness, launched in beta in August 2026 alongside the Muse Spark 1.2 model. It is the same category of tool as Claude Code, Codex CLI, and OpenCode — you run it in your terminal, it reads your project files, edits code, and runs commands.
The harness is built specifically for Muse Spark models but supports the same agentic patterns developers expect in 2026: a /goal loop for autonomous task execution, sub-agents for parallel work, skills for reusable workflows, and adjustable reasoning effort levels. Meta has not open-sourced the harness itself — Mark Zuckerberg confirmed this in the launch announcement (Meta Research Blog).
The installer is a 97 MB download, which suggests a Node.js-based application (the Node engine alone is ~80-90 MB). Several major coding harnesses in 2026 share this characteristic — only a few, like Codex CLI (Rust) and Grok Build (Rust), are compiled to native binaries.
What Is Muse Spark 1.2?
Muse Spark 1.2 is Meta's coding-focused language model with a 1-million-token context window. Meta describes it as "optimized for real coding workflows, with higher first-attempt accuracy and more reliable tool calling" (Meta Developer Portal).
Meta reports benchmark scores on Terminal-Bench, DeepSWE, Meta Internal Code Bench, and GDPVal, though the developer page displays these as charts rather than publishing specific numbers. The model is the successor to Muse Spark 1.1, which launched earlier in 2026 and was described as "decent" but not frontier-class.
Muse Spark 1.2 is available through three channels:
- Meta Model API — direct, self-serve access from Meta
- OpenRouter — swap it into existing tools and workflows (OpenRouter listing)
- Muse Code — Meta's own harness, built for the model
How Much Does Muse Spark 1.2 Cost?
The pricing is the most interesting part of this launch. Meta offers two tiers, and the difference is enormous:
| Tier | Context | Input (per 1M tokens) | Cached input (per 1M) | Output (per 1M) | Data used for training? |
|---|---|---|---|---|---|
| Contributor | 1M | $0.10 | $0.002 | $0.20 | Yes |
| Non-contributor | 1M | $1.25 | $0.15 | $4.25 | No |
Source: Meta Developer Portal — Muse Spark 1.2 pricing, verified August 6, 2026.
The contributor tier is 12.5x cheaper on input tokens and 21x cheaper on output. But the cached input rate is where it gets extreme: $0.002 per million cached tokens (contributor) vs $0.15 (non-contributor) — a 75x difference.
Why cached pricing matters for coding agents
In any long-running agent session — where you give the model a goal and it works for 20, 30, or 40 minutes — the majority of your token consumption is cached input. Every turn, the model needs to see its own previous response to decide the next step. That response becomes cached input on the next call.
The presenter's demo burned about 1 million tokens in a single session building a platformer game. Over 85% of those tokens were cached input. At the contributor tier, 1 million cached input tokens cost $0.002 — effectively free. At the non-contributor tier, the same cached tokens cost $0.15. At Claude Opus-level pricing, the same session could cost $2-4.
This makes the contributor tier particularly attractive for:
- Personal projects and experiments
- Open-source work where code is already public
- Learning and prototyping
But do not use contributor mode for company code. If you are working on proprietary codebase, the non-contributor tier ($1.25/$0.15/$4.25) is the only safe option, and the price advantage shrinks dramatically. For a deeper look at cost optimization across providers, see our guide on cutting AI inference costs with open-source model routing.
Muse Code Harness Features
Voice input
Muse Code has voice input built in. You press a key, speak your prompt, and the harness transcribes it — with light post-processing (it added a period to the end of a spoken prompt in the demo). This works over the API.
Claude Code, by comparison, only offers voice input with a Claude subscription — not over the API. If you build custom tools on top of an API and want voice, Muse Code is one of the few harnesses that delivers it natively.
Sub-agents and skills
The harness supports sub-agents (fan-out workers for parallel tasks) and skills (reusable workflows). Sub-agents can be configured to run in auto or off mode. Skills work the same way as in other harnesses — you define a workflow once and invoke it repeatedly.
For a deeper architectural look at how sub-agents and durable logs make agents reliable, see our analysis of AI agent architecture and why the log is the system.
Effort levels
Muse Code offers six reasoning effort levels: ultra, extra high, high, medium, low, and minimal. Higher effort means deeper reasoning but slower responses. The presenter used "extra high" for testing the model's ceiling.
Web search
Web search has three modes: client (searches through your computer's network), hosted (uses Meta's infrastructure), and off. Client is the default. This is a useful distinction — if you are behind a corporate firewall or need search results from your own network context, client mode handles it.
Session messaging
One feature that sets Muse Code apart: session messaging — the ability for local sessions to message each other. This would let you run multiple agents on related tasks and have them coordinate, rather than one agent and its sub-agents working in isolation. The presenter noted this was not fully enabled in the current beta ("the local session messaging gate is closed"), but it is a feature no other major harness offers.
This matters for workflows where multiple top-level agents work on related branches and need to avoid overwriting each other's changes — a pattern that sub-agents alone do not solve well.
Other features
/goalloop — autonomous task execution (same as Codex CLI's/goal)/side— side conversation mode (similar to Claude Code's/btw)- YOLO mode — auto-approve all actions (restart required to enable)
muse resume— resume a previous session from a fresh terminal- Themes — customizable UI, including warm dark mode options
- Agent delegation — auto or off, controls whether a helper model handles delegation
- Summaries — on/off toggle for session summaries
The Demo: Building a Platformer Game with Voice
The presenter tested Muse Code by speaking a voice prompt: "I want to create a web-based game, a platformer game." The harness:
- Transcribed the voice prompt with light post-processing
- Asked clarifying questions about creative direction
- Built a working platformer game called "Skybound" with momentum physics, jump mechanics (spacebar), and double-jump
- Started a static server so the game was immediately playable
The game was described as "decent" — comparable to or better than what AI models produced 3-4 months earlier in similar tests. It included gravity mechanics, directional movement (arrow keys), and a double-jump booster.
The session consumed approximately 1 million tokens, with over 85% as cached input. At the contributor tier, this cost roughly $0.02. At the non-contributor tier, the same session would have cost $0.20-0.30.
How Does Muse Spark 1.2 Compare to Other Models?
The presenter was clear: Muse Spark 1.2 is not a frontier-tier model. It is not as powerful as Claude Fable or GPT-5.6 Sol at extra-high effort. But it is a capable coding model, and the pricing makes it competitive for specific use cases:
| Model | Input (per 1M) | Cached input (per 1M) | Output (per 1M) | Context | Open weights? |
|---|---|---|---|---|---|
| Muse Spark 1.2 (contributor) | $0.10 | $0.002 | $0.20 | 1M | No |
| Muse Spark 1.2 (non-contributor) | $1.25 | $0.15 | $4.25 | 1M | No |
| DeepSeek V4 Flash | $0.14 | $0.0028 | $0.28 | 1M | Yes (MIT) |
| Claude Opus 4.8 | ~$15.00 | ~$1.50 | ~$75.00 | 200K | No |
Sources: Meta Developer Portal, DeepSeek API pricing, Anthropic API pricing page. Verified August 2026.
At the contributor tier, Muse Spark 1.2 is priced similarly to DeepSeek V4 Flash ($0.14/$0.28) — and both are dramatically cheaper than Claude or GPT-5.6. The difference is that DeepSeek V4 Flash has open weights (MIT-licensed), while Muse Spark does not.
If you want an open-source, model-agnostic harness that works with any of these models, OpenCode supports 75+ providers and can swap in Muse Spark via OpenRouter. For a fully open-source agent setup, see our guide on building a free AI agent with Hermes Agent and Kimi K3.
Meta also publishes cookbooks for Muse Spark including multi-agent orchestration (standing up a four-profile agentic team), agent fan-out (spinning subagents into isolated worktrees), computer use, and a GitHub agent built with Muse Spark and OpenCode (Meta Cookbook).
Should You Use Muse Code?
Use it if:
- You want the cheapest capable coding agent for personal or open-source projects
- You are OK with Meta training on your code (contributor tier)
- You want voice input in a terminal coding agent
- You want to try session messaging between parallel agents (once it leaves beta)
Skip it if:
- You work on proprietary company code — the non-contributor tier is not price-competitive with DeepSeek or local models
- You need an open-source harness you can audit and fork
- You need frontier-tier code quality (use Claude Opus or GPT-5.6 Sol instead)
The honest take
Meta has a track record of building developer products — React and React Native transformed web and mobile development. They know how to build developer tools. Muse Code is a solid first harness, and the pricing is genuinely disruptive for the contributor tier. But the model itself is mid-tier, and the harness is not open source. If Meta invests in this product the way they invested in React, it could become a serious contender. Today, it is a cheap, interesting beta worth trying on non-sensitive work.
What This Means for You
If you are a developer evaluating coding agents: try Muse Code with the contributor tier on a personal project to understand its capability ceiling. The cost is negligible — a full session building a game cost $0.02. Use it as a data point when comparing against Claude Code, Codex CLI, and OpenCode.
If you are a team lead or CTO: the non-contributor tier pricing ($1.25/$4.25) is not competitive enough to justify switching from your current stack. The contributor tier is off-limits for proprietary code. Watch this space if Meta open-sources the harness or improves the model.
If you are a builder or small-business owner: if you have a non-sensitive project — a marketing site, an open-source tool, a prototype — the contributor tier is one of the cheapest ways to get an agentic coding session running. Pair it with a free AI code review tool to catch bugs, since the model is not frontier-tier.
FAQ
Q: Is Muse Code free to use?
A: No. Muse Code is in beta and requires a payment method. The Muse Spark 1.2 model has two pricing tiers: contributor ($0.10/M input, $0.002/M cached, $0.20/M output) and non-contributor ($1.25/$0.15/$4.25). The contributor tier is extremely cheap but allows Meta to train on your data. Source: Meta Developer Portal.
Q: Is Muse Code open source?
A: No. Mark Zuckerberg confirmed the harness is not open source in the launch announcement. The model (Muse Spark 1.2) is also not open weight. If you need an open-source coding agent, consider OpenCode (160K+ GitHub stars, 75+ model providers) or Codex CLI (Apache 2.0). Source: Meta Research Blog.
Q: How does Muse Spark 1.2 compare to Claude Opus or GPT-5.6?
A: Muse Spark 1.2 is not a frontier-tier model. It is a coding-focused model that Meta says is "competitive with the best" but the presenter's testing found it less capable than Claude Fable or GPT-5.6 Sol at high effort. Its advantage is price, not capability. Source: presenter testing, Meta Developer Portal.
Q: Can I use Muse Spark 1.2 with other harnesses?
A: Yes. Muse Spark 1.2 is available on OpenRouter, so you can swap it into any OpenAI-compatible harness including OpenCode, Cursor, or your own tooling. Source: OpenRouter listing.
Q: What is the contributor tier and should I use it?
A: The contributor tier means Meta can use your prompts and code to improve their products. In exchange, you get pricing that is 12-75x cheaper than the non-contributor tier. Use it for personal projects, open-source work, or anything you are comfortable sharing with Meta. Do not use it for proprietary company code. Source: Meta Developer Portal.
Q: Does Muse Code have voice input?
A: Yes. Voice input is built into the harness and works over the API. Claude Code only offers voice input with a subscription, not over the API, making this a unique feature among terminal coding agents. Source: presenter demo, Meta Research Blog.
Discussion
0 comments