Meta Muse Spark 1.2 & Muse Code: Terminal Agent With Parallel Sub-Agents and $0.10/M Contributor Pricing

A developer at Meta's internal hackathon reportedly handed Muse Code a 200,000-line monorepo, told it to migrate a deprecated authentication library across every service, and walked away. Four hours later — without the developer touching the keyboard — Muse Code had opened twelve parallel sub-agent sessions, each working inside its own isolated git worktree to avoid conflicts, and committed a complete set of changes including updated unit tests. That story, circulating on developer forums since the August 5 release, captures what makes this launch different from the string of AI coding tools that have come before it: the parallelism isn't a configuration option, it's the architecture.

Vitals

Model: Muse Spark 1.2
Context window: 1,000,000 tokens
Modalities: Text, image, video, audio, PDF (input); text (output)
License: API — closed (open weights announced but not yet released at launch)
Standard pricing: $1.25 input / $4.25 output per million tokens
Contributor pricing: $0.10 input / $0.20 output per million tokens (data opt-in required)
Release date: August 5, 2026 (Muse Code beta + Spark 1.2 API; open weights TBD)
Availability: macOS, Linux terminal install; API via Meta AI platform

How the agent architecture works

Muse Code is a local-first terminal agent that wraps Muse Spark 1.2. Unlike editor plugins or chat-based coding assistants, it runs as a persistent background process with an append-only event log — meaning a session survives crashes, disconnections, and system restarts without losing its place. The agent can be started, paused, and resumed at will.

The key architectural decision is multi-agent parallelism via isolated git worktrees. When Muse Code encounters a task that can be decomposed — migrating a library, refactoring a module across multiple services, writing test suites in parallel with implementation — it spawns multiple sub-agents. Each sub-agent works in its own worktree, a Git mechanism that creates a separate working directory for the same repository without switching branches in the main checkout. This prevents file conflicts and lets the orchestrator merge results after independent completion. The coordination layer is built into Muse Code itself: the orchestrator assigns tasks, monitors sub-agent progress, and handles merge resolution.

Built-in slash commands give developers approval gates without leaving the terminal: /plan generates and stages a plan for review before execution begins, /grill stress-tests the plan by probing for failure modes, and /goal sets a persistent objective the agent pursues across sessions.

Benchmarks and positioning

Meta has not published a full standalone benchmark suite for Spark 1.2 comparable to a Frontier-Bench or SWE-bench disclosure. The official release materials emphasize real-world task completion — whole-repository generation, multi-file refactoring, debugging across large codebases — over synthetic benchmark scores. That's a deliberate positioning choice: by making Muse Code the demo artifact rather than a number, Meta is betting on hands-on reception over benchmark marketing. Independent evaluations are expected in the coming weeks as developers get access to the beta.

What is verified: Muse Spark 1.2 supports a 1-million-token context window, meaning the model can ingest an entire mid-size codebase in a single pass without chunking or retrieval. That distinguishes it from agents built on models with shorter context windows that rely on retrieval-augmented generation (RAG) to manage large repositories. Meta has positioned this as "context-native" development — no RAG pipeline, no chunk management, no lost context.

The Contributor tier is the real story for indie developers

The pricing structure is what makes this release strategically notable beyond the architecture. Standard API access — where Meta does not use prompts or completions for training — costs $1.25/$4.25 per million tokens, competitive with comparable frontier API pricing. But the Contributor tier prices at $0.10 input / $0.20 output per million tokens, a 92% discount, in exchange for explicit opt-in to allow Meta to use your sessions as training data.

This is a direct acknowledgment of the data flywheel problem: training frontier coding agents requires large quantities of real coding trajectories, which are expensive and difficult to source. Meta is essentially offering subsidized compute in exchange for data generation. For individual developers running intensive agentic sessions, the discount changes the economics entirely — a long Muse Code session that might cost $30–40 at standard rates drops to under $3 at Contributor rates. The tradeoff is explicit and clearly disclosed, unlike some historical cases where training data use was buried in terms of service.

Try it yourself

These are illustrative prompts based on Muse Code's reported capabilities — not prompts verified by this publication's own testing.

muse-code run --goal "Migrate all uses of deprecated AuthV1 library to AuthV2 across the entire repository, update unit tests, and open a draft PR when complete"
muse-code plan "Add input validation and error handling to all public API endpoints in /src/api/ — generate a plan for my review before making any changes"
muse-code run "Find all instances where database queries are made outside of transaction blocks and refactor them to use the existing transaction utility"

What this means for the agent space

The Muse Code release arrives in a market crowded with AI coding tools, but most of them are editor integrations or single-agent completion assistants. Persistent multi-agent coordination inside git worktrees is a qualitatively different capability — it moves the model from "sophisticated autocomplete" to something closer to an asynchronous engineering collaborator. Whether the real-world reliability lives up to the architecture's ambition will be the key question as the beta opens to wider access.

For pricing, the Contributor tier is a structural challenge to the market: it's hard for competing tools to match $0.10 input pricing without a comparable data strategy. Meta's open-source model heritage — and its track record of releasing weights for prior Muse Glimmer (released August 10) — suggests the eventual open-weight release of Spark 1.2 will add another pressure point on closed-model coding tools.

Frequently Asked Questions

What is Meta Muse Code?

Muse Code is a terminal-native AI coding agent released by Meta on August 5, 2026. It runs persistent background agents that manage complex, multi-step software engineering tasks using Muse Spark 1.2 as the underlying model. It can spawn parallel sub-agents in isolated git worktrees to handle large repository tasks without file conflicts.

What is the Contributor pricing tier for Muse Spark 1.2?

The Contributor tier prices at $0.10 input / $0.20 output per million tokens — approximately 92% cheaper than the standard $1.25/$4.25 pricing. In exchange, users must opt in to allowing Meta to use their prompts and completions as training data. The opt-in is explicit and disclosed upfront.

Does Muse Code use RAG?

No. Muse Code is designed around Muse Spark 1.2's 1-million-token context window, allowing it to ingest entire codebases in a single pass. Meta positions this as 'context-native' development that avoids the complexity and context-loss risks of retrieval-augmented generation pipelines.

When will Muse Spark 1.2 open weights be released?

Meta announced plans to release open weights for Muse Spark 1.2 but had not published them at the time of the August 5 launch. A release timeline has not been officially confirmed.

Meta's Muse Code Is a Terminal Agent That Runs Parallel Sub-Agents in Isolated Git Worktrees