Meta Muse
Meta's terminal-based AI coding agent — Muse Code on the Muse Spark reasoning model, with data-for-discount pricing that undercuts rivals up to 20×.
1. Core Product / Service
Muse is Meta's first entry into the AI coding-agent market, launched by Mark Zuckerberg into public beta on August 5–6, 2026 [2]. The flagship product, Muse Code, is a command-line agent (the muse command, installable via a one-liner from dev.meta.ai) that runs on macOS and Linux and handles end-to-end software-engineering tasks across large repositories — planning changes, writing code, and validating results [2][3]. It is powered by the Muse Spark 1.2 coding model, which supports up to 1 million tokens of context for reasoning over thousands of files at once [2].
Signature capabilities: persistent async background agents that work in parallel on different parts of a project, each inside an isolated workspace; a local event log that records every model call, tool action, approval, and edit so the runtime is restart-safe; and built-in commands including /plan (an approval-gated plan), /grill (adversarially stress-test the plan), and /goal (run until the objective is done) [2]. The isolation-plus-gating design — isolated workspaces with human-approval gates on plans — is the mechanism Jimmy flagged as Muse's differentiator versus manus, hermes-agent and OpenClaw [1].
On September 1, 2026 Muse Code exited beta with inter-session messaging, a "Workflow" mode for coordinating many sub-agents, a "Rewind" function, an SDK developer preview, and subscription plans [5]. Muse Code and the Meta Model API are also distributed through the Oracle Cloud Marketplace [6].
2. Target Users & Pain Points
- Developers on large, long-horizon codebases who want an agent that keeps working across sessions and survives crashes — served by the persistent agents + restart-safe event log.
- Cost-sensitive individual developers / startups willing to trade training-data rights for a ~20× price cut (Contributor tier) [4].
Pain solved: single-shot coding assistants fall over on multi-hour, multi-file tasks. Muse's bet is that background sub-agents in isolated workspaces, plus explicit approval gates, produce the reliability that a "fire and check back later" workflow needs [1][2].
3. Competitive Landscape
| Tool | Vendor | Model | Terminal-Bench 2.1 |
|---|---|---|---|
| Muse Code | Meta | Muse Spark 1.2 | 82.9% |
| Claude Code | anthropic | Opus 5 | 86.7% |
| Codex | openai | GPT-5.6 Terra | 81.8% |
| Grok Code | xAI | Grok 4.5 | 81.6% |
| github-copilot | Microsoft/GitHub | — | — |
On Terminal-Bench 2.1, Muse Spark 1.2 scored 82.9% — behind Claude Opus 5 (86.7%) but ahead of GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%); on DeepSWE 1.1 it scored 59.3% (third), and on Meta's internal benchmark 70.6% versus Opus 5's 79.4% [2]. Meta emphasizes long-horizon reliability (a 1,000+-tool-call, up-to-24-hour GPU-kernel optimization case study) over raw benchmark wins [2].
4. Unique Observations
- "Data-for-discount" pricing is the real story. The Standard tier costs $1.25/M input and $4.25/M output; the Contributor tier drops to $0.10/M in / $0.20/M out — up to 20× cheaper — in exchange for letting Meta train on your code [4]. This is Meta's advertising-company economics transplanted to agents: subsidize inference with training data, the same way ads subsidize consumer products. OpenAI and Anthropic charge a premium for not training; Meta charges a discount for letting it train [4].
- Open-to-closed pivot, concretely. Muse Spark is closed-weight and API-only — the product-level continuation of Meta's shift away from open-weight Llama identity, per meta-superintelligence-labs [2].
- Local comparison set. Jimmy's 2026-10-02 notes compare Muse Code against ZCode / OpenCode / DeepOpenCode (local CLI tools) and Muse.ai against hermes-agent / OpenClaw / manus — positioning Muse as an agent-stack competitor, not just an IDE plugin [1].
5. Financials / Funding
Muse is a Meta product line (no external funding). Pricing [4][5]:
- API: Standard $1.25/M in / $4.25/M out; Contributor $0.10/M in / $0.20/M out (training-data opt-in); zero-data-retention available on request.
- Subscriptions (post-beta): $5/mo Everyday, $15/mo High Usage, $50/mo Power Usage.
6. People & Relationships
- Announced by: Mark Zuckerberg (CEO, Meta) [2].
- Org: sits under meta-superintelligence-labs (Meta's consolidated AI unit), which shipped the Muse Spark reasoning-model line.
- Competitors: anthropic (Claude Code / Opus 5), openai (Codex), GitHub / github-copilot, cursor, windsurf, devin.
- Distribution: Oracle Cloud Marketplace [6].
Sources
[1] local: 2026-10-02-summary.md — Muse Code vs ZCode/OpenCode; Muse.ai vs Hermes Agent/OpenClaw/Manus; isolation + gating mechanism; "data-for-discount" strategy. [2] VentureBeat, "Meta enters the AI coding wars with Muse Spark 1.2 and Muse Code," https://venturebeat.com/orchestration/meta-enters-the-ai-coding-wars-with-muse-spark-1-2-and-muse-code-with-persistent-async-background-agents (2026-10-05) [3] The Next Web, "Meta enters the AI coding wars with Muse Code and a price to match," https://thenextweb.com/news/meta-muse-code-ai-coding-agent-launch (2026-10-05) [4] MacObserver, "Meta Wants Your Coding Data, and It'll Cut Muse Code Prices by Up to 20x," https://www.macobserver.com/news/meta-wants-your-coding-data-and-itll-cut-muse-code-prices-by-up-to-20x/ (2026-10-05) [5] ITHome, "Meta AI 编程工具 Muse Code 结束测试，新增多项功能并开启订阅服务," https://m.ithome.com/html/996937.htm (2026-10-05) [6] Oracle, "Meta Model API and Muse Code on Oracle Marketplace," https://blogs.oracle.com/cloud-infrastructure/access-meta-through-oracle-marketplace (2026-10-05)