---
title: "Meta's Muse Code Undercuts Claude Code by 21x — If You Let Meta Train on Your Repo"
date: 2026-08-08
tags: ["meta","muse-code","muse-spark","agentic-coding","cli","pricing","benchmarks"]
categories: ["AI Tools","Industry"]
summary: "Meta shipped Muse Code, a terminal coding agent powered by Muse Spark 1.2, on August 5. It has persistent background agents, an append-only event log, and a price list with two columns: $1.25/$4.25 per million tokens, or $0.10/$0.20 if you grant Meta the right to train on your prompts and completions. The second column is the actual product."
---


![A balance scale weighing a single coin against a folder of source code being carried away](/images/meta-muse-code-terminal-agent-contributor-tier.png)

Meta shipped a terminal coding agent on August 5. It's called Muse Code, it's in beta, it installs with one curl command, and it runs on a new coding-tuned model, Muse Spark 1.2. That's the announcement. The interesting part isn't in the announcement — it's in the pricing table.

Muse Code costs $1.25 per million input tokens and $4.25 per million output. Or it costs $0.10 and $0.20 for the *identical model*, if you sign up for what Meta calls the Contributor tier and grant permission to use your prompts and completions to train future Meta models. That's 12.5x cheaper on input, 21x cheaper on output, and 75x cheaper on cached input ($0.002 versus $0.15).

Meta is not selling you a coding agent. Meta is buying your codebase, and the agent is what it's paying with.

## What Actually Shipped

Credit where it's due: Muse Code is a real terminal agent, not a chat window with a shell attached. The architecture choices are the ones you make when you've watched what Claude Code did and decided to compete on the same terrain rather than retreat into an IDE.

Three things stand out from Meta's own description:

**Persistent async background agents.** Subagents stay alive across the session rather than being spawned, used, and discarded per task. This is the delegation model Claude Code arrived at with agent teams, and Meta went straight to it rather than shipping a single-threaded loop first.

**A local append-only event log.** Every model call, tool run, approval, and edit gets appended to one log that serves as the runtime's single source of truth. Meta's framing is that this makes the runtime "replay-exact and restart-safe." This is the correct design and it's genuinely underrated — an agent whose entire history is a replayable log is an agent you can audit, resume after a crash, and reason about after the fact. Most agent runtimes treat state as an implementation detail. Treating it as the product is a good instinct.

**Bundled skills.** `/plan` turns a task into an approval-gated plan, `/grill` stress-tests that plan until it holds up, `/goal` works toward an objective end to end. The `/grill` idea — an adversarial pass over your own plan before execution — is the one worth stealing. Anyone running spec-driven workflows already knows the failure mode it targets: a plan that reads well, gets approved, and turns out to be built on an assumption nobody challenged.

Install is `curl -fsSL https://dev.meta.ai/install.sh | bash` on macOS and Linux. Piping a vendor's install script straight into bash is its own conversation, and it's the same conversation every terminal agent has forced; at least the pattern is now consistent enough that most teams have a policy about it.

Muse Spark 1.2 itself is the coding-focused successor to July's 1.1. Meta says it scaled up training compute on coding tasks, expanded environment diversity, and — the notable detail — co-trained the model with Muse Code itself using rejection-sampled trajectories from the agent. It also used Muse Spark 1.1 to generate training data for 1.2's instruction-following. Model and harness trained together, with the previous model producing the data. The showcase result is a GPU kernel optimization run that held together across 1,000+ tool calls.

## The Numbers, and Who Checked Them

Meta's self-reported Terminal-Bench 2.1 score for Muse Spark 1.2 is **82.9**, against **86.7** for Claude Opus 5. Meta reports 1.2 in second place behind Opus 5 on all three coding benchmarks it published — Terminal-Bench 2.1, DeepSWE v1.1, and Meta's internal coding bench. To Meta's credit, the launch materials carry an explicit caveat that its evaluation setup "may not be specifically tuned for proprietary third-party models," so competitor numbers may not reflect those models at their best.

This blog has a history with Meta's benchmark numbers. When Muse Spark 1.1 launched in July claiming 80.0 on Terminal-Bench 2.1, Vals AI's independent rerun measured 69.29 — [an eleven-point gap](/2026/07/meta-muse-spark-1-1-paid-api-benchmark-gap/) that turned "competitive with Opus 4.8" into "trailing badly."

So the honest thing to report is that this time it held up better. Artificial Analysis, running its own harness on launch day, measured Terminal-Bench v2.1 at **80%** against Meta's claimed 82.9 — a roughly three-point gap, which is ordinary harness variance rather than a credibility event. Artificial Analysis puts Muse Spark 1.2 at **54** on its Intelligence Index, behind Claude Opus 5 at **61** and GPT-5.6 Sol at **59**, at $0.40 per Intelligence Index task.

That's a coherent picture, and it's not an embarrassing one: a solid second-tier coding model, a few points behind the frontier on the benchmarks that matter for agentic work, at a fraction of frontier pricing. Meta's self-reporting was closer to reproducible this round than last. Neither 1.2 nor Opus 5 has an independently verified entry on the official Terminal-Bench leaderboard yet, so treat the exact ranking as provisional.

## The Price Has Two Columns

Now the part that matters.

| | Standard | Contributor |
|---|---|---|
| Input / 1M | $1.25 | $0.10 |
| Cached input / 1M | $0.15 | $0.002 |
| Output / 1M | $4.25 | $0.20 |
| Your data | Not used to improve Meta's products | Used to improve Meta's products |
| Rate limits | 3,000 req/min, 4M tokens/min | 60 req/min, 2.1M tokens/min |

Look at the rate limits, because they tell you what Meta thinks it's buying. The cheap tier is throttled to **1/50th the request rate** of the expensive one. This is not a plan designed for a company to run its production agent fleet on. It's designed for individual developers working at human pace on real repositories — which is precisely the data that's expensive to synthesize and impossible to scrape: intent, iteration, failed attempts, and the corrections that follow.

Meta already has the world's code from GitHub scrapes. What it doesn't have is *how developers actually get to* that code — the trajectory, not the destination. Muse Code's event log captures exactly that, and the Contributor tier is the mechanism for shipping it home. The co-training detail from the model announcement is the strategy stated in the open: agent and model improve together on trajectories, and the Contributor tier is the trajectory pipeline.

Sources disagree on one operational detail worth nailing down before you install anything: at least one outlet reports Muse Code defaults to the Contributor tier after install, while others state plainly that Standard is the default and Contributor requires an explicit signup through a direct Meta relationship (it isn't offered through aggregators like OpenRouter). Meta's published pricing table labels the tiers "Used to improve our products" versus "Not used to improve our products," which supports the opt-in reading. **Check your tier before your first real prompt anyway.** The cost of being wrong is asymmetric and unrecoverable.

And be clear about who can actually take this deal. If you're on client work under an NDA, the permission to train on that code isn't yours to grant. If your repo carries proprietary IP, the savings are a rounding error against what you're handing over. The tier is genuinely fine for personal projects, open-source work, throwaway prototypes, and public-code experimentation — that's a real segment, and $0.10 per million tokens is a real gift to it.

## What This Means for Your Stack

Muse Code doesn't change the frontier. On Meta's own numbers it's second to Opus 5 across the board, and Artificial Analysis's independent index agrees. If your bottleneck is the hardest 10% of engineering work — the gnarly refactor, the subtle concurrency bug, the migration that touches 400 files — nothing here argues for moving off Claude Code, and the architecture that makes Claude Code work (terminal-native, genuinely autonomous, auditable) is the architecture Meta just validated by copying it.

Where Muse Code is worth an hour of your time: high-volume, low-stakes agent work on code you don't mind sharing. Bulk documentation passes. Lint-and-fix sweeps across an open-source monorepo. Dependency bumps. The same budget-tier niche GLM-5.2 and Kimi K2.7-Code occupy among open-weight models — except this one is closed and proprietary, so you get neither the price of self-hosting nor the control.

The genuinely portable idea is `/grill`. An adversarial pass that attacks your own plan before a single file gets written is a good enough pattern that you should implement it in whatever agent you already run, today, for free. That's a `CLAUDE.md` instruction and a subagent, not a product migration.

And the broader signal is worth naming: the price of a competent coding agent is now openly denominated in your source code. Meta put both columns on the same page and let you pick, which is more honest than tools that quietly train on your telemetry without a discount. But an honest price tag is still a price tag. Read it before you pipe the install script into bash.

---

**Sources:**
- [Meta AI Research — Introducing Muse Code and Muse Spark 1.2](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2)
- [Meta AI Developers — Meet Muse Spark 1.2 and Muse Code](https://developer.meta.com/ai/resources/blog/build-with-muse-code/)
- [Artificial Analysis — Muse Spark 1.2](https://artificialanalysis.ai/articles/muse-spark-1-2)
- [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)
- [Engadget — Meta introduces Muse Code, its take on a coding agent](https://www.engadget.com/2231285/meta-introduces-muse-code-its-take-on-a-coding-agent/)
- [MacRumors — Meta's new Mac coding agent costs up to 20x less if you let Meta train on your data](https://www.macrumors.com/2026/08/05/meta-muse-code-for-mac/)
- [Digital Applied — Muse's Contributor Tier: Cheap Tokens, Real Trade-Offs](https://www.digitalapplied.com/blog/muse-spark-contributor-tier-economics-tradeoffs)