Meta Muse Code Review (2026): The $0.20/M Contributor Tier, Real Capabilities, and When to Use It

TL;DR

Muse Code is Meta's first coding agent, launched in public beta on August 5, 2026 alongside the Muse Spark 1.2 model. It runs in your terminal on macOS or Linux, plans multi-file changes, writes code, and validates the results — the same broad workflow as Claude Code or Codex.

The story is pricing: a Contributor tier at $0.10 input / $0.20 output per million tokens, roughly 21x cheaper than the standard tier — and roughly 50x cheaper than Claude Sonnet 5 on output tokens. The catch is data: the Contributor tier explicitly trains on your prompts, code, and completions.

Three things are actually new versus Claude Code / Codex: persistent sub-agents that don't lose context between subtasks, a crash-safe local event log for resume-after-failure, and bundled /plan, /grill, and /goal skills.

Benchmarks are vendor-reported, not independently verified. Treat them as Meta's starting point, not settled fact.

What is Muse Code?

Muse Code is a terminal-based coding agent built by Meta Superintelligence Labs, led by Alexandr Wang. It runs from the command line on macOS and Linux, takes on complex software-engineering tasks across large repositories, and can plan changes, write code, and validate results autonomously.

Unlike Cursor or Windsurf, Muse Code has no GUI, no VS Code extension, and no JetBrains plugin. It is a pure terminal agent, in the same form factor as Claude Code and Codex CLI.

The underlying model is Muse Spark 1.2, Meta's coding-focused model co-trained with the Muse Code harness from day one. Meta's argument is that co-training produces more reliable tool use and fewer wasted retries. That claim is hard to verify independently — we have no benchmark from a third party yet.

Muse Spark 1.2 also carries a 1 million-token context window, inherited from the Muse Spark line. For comparison, that matches Opus 5 and Sonnet 5.

Install & first run

Installation is a single shell command on macOS or Linux:

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

Windows is not currently supported. If you're on Windows, use WSL.

After the installer finishes, run muse-code from any directory. It opens a browser-based auth flow against your Meta developer account. Once authenticated, the agent is ready.

First-run tips:

• Start in a real project directory. cd ~/your-project first, then launch. Muse Code reads your file structure to understand context.
• Pick your tier early. Switch via /model muse-spark-1.2-contributor or /model muse-spark-1.2 inside the agent. The Contributor tier is rate-limited by tokens in a rolling 5-hour window, not by request count — useful for day-to-day use.
• For long-running tasks, run on a machine that won't sleep. Long jobs (1000+ tool calls) are exactly what the agent is designed for, but laptop sleep kills them.

Pricing: two tiers explained

Muse Code is pay-per-token only — no subscription, no monthly minimum. The two tiers differ by both price and data policy:

To put the Contributor tier in context:

The Contributor tier is roughly 50x cheaper on output than Claude Sonnet 5, and beats every frontier closed model on price. But the data tradeoff is real: Meta explicitly states that Contributor-tier prompts and completions feed into training future models. If you work on proprietary code, trade secrets, or anything covered by an NDA, this tier is off-limits. Use the standard tier or contact Meta sales for zero-data retention.

For open-source work, personal projects, or code you don't mind Meta seeing, the economics are compelling.

Core capabilities

Five things differentiate Muse Code from a generic Claude Code / Codex clone:

1. Persistent sub-agents

Muse Code does not spawn fresh agents for each subtask. Instead, persistent background agents remain active throughout the session, carry context forward, and decide when to report back to the main agent. The benefit shows up on multi-step refactors where other agents lose context mid-task and need re-steering.

2. Crash-safe runtime

Every model call, tool run, approval, and file edit is appended to a local event log. The runtime is replay-exact and restart-safe. If your laptop sleeps or the process is killed, Muse Code resumes from where it stopped — no re-deriving context. For jobs that run hundreds or thousands of tool calls (kernel optimization, large migrations), this is closer to a requirement than a nice-to-have.

3. Bundled skills

Muse Code ships with built-in slash commands you can call from any task:

• /plan — turn a task into an approval-gated plan before execution.
• /grill — stress-test that plan until it holds up under adversarial questioning.
• /goal — drive a multi-step task toward successful completion of a stated objective.

4. Multimodal input

Muse Code accepts video and image input. In Meta's launch demo, the agent interpreted a fly-through video of a home and produced a visually rich marketing page from it. The practical implication: you can pipe screenshots, screen recordings, or mockup images directly into a task.

5. Long-horizon scale

Meta demonstrated the agent making over 1,000 tool calls in a single 24-hour GPU kernel optimization task, progressively improving NVIDIA Hopper kernel performance. That's not a benchmark most users will hit, but it's the workload the harness was co-trained on.

Benchmarks: reality check

Muse Spark 1.2's headline numbers, from Meta's own research blog:

On Meta's own comparison charts, Muse Spark 1.2 trails Claude Opus 5 on every coding benchmark shown, while beating GPT-5.6 Terra and Gemini 3.6 Flash on most. The 82.9% TerminalBench score is a 6.7-point improvement over Muse Spark 1.1, released a month earlier.

Every number is vendor-reported. Each model was tested inside its own harness. There is no neutral third-party audit of these results yet. Treat the table as Meta's starting position, not settled fact. The pricing story is much more durable than the benchmark story.

vs Claude Code & Codex

All three are terminal coding agents. Where they differ is the harness around the model:

The honest summary: Claude Code and Codex are still the capability leaders, Muse Code is the cost leader and the architectural innovator (sub-agents, crash safety, multimodal). The right pick depends on what you're optimizing for.

Who should try it

Try Muse Code if:

• You run high-volume, repetitive agent tasks — batch refactors, migrations, test generation, dependency upgrades — where token cost matters more than peak intelligence.
• You work on open-source code and don't mind Meta seeing your prompts and completions (Contributor tier).
• You want a second-opinion agent at low cost, and don't need it to be the only one you run.
• You value crash-safe long-horizon runs for multi-hour migrations or kernel-style optimization.

Hold off if:

• You need frontier reasoning for complex architectural work — Meta openly concedes Muse Code isn't leading on raw capability. Opus 5 / Sonnet 5 / GPT-5.6 Sol score higher on vendor benchmarks.
• You work on proprietary, NDA-covered, or trade-secret code. The Contributor tier trains on your code; the standard tier is fine but loses the cost advantage.
• You're on Windows without WSL. Muse Code does not support Windows natively today.
• You rely on IDE-native features (VS Code extension, inline diff review, file tree in agent). Muse Code is pure terminal.

Muse Code is a beta product. Treat it as such: test on non-critical repos, compare cost-per-task against your current tool, and watch whether the Contributor tier survives the beta. If the async sub-agent architecture delivers the latency wins Meta claims, the incumbents will need to respond on price — which is good news for everyone's bill.

How we wrote this

This piece was compiled on August 7, 2026, two days after Muse Code's public beta launch. We cross-referenced:

• Meta's official launch blog and the research post.
• Third-party technical coverage from Mayhem Code, AI Made Tools, and AppleInsider.
• Comparison pricing for Claude Sonnet 5 / Opus 5 / GPT-5.6 / DeepSeek V4-Flash from our August 2026 model pricing tracker.

We have not independently installed or tested Muse Code. Benchmarks are vendor-reported. Pricing is current as of August 7, 2026 — Claude Sonnet 5 introductory rates expire August 31, and Contributor-tier availability / pricing may shift during the beta.

Corrections or counter-evidence welcome in the comments or via our about page contact.