Muse Code and Muse Spark 1.2: Meta's First Real Coding Agent, Fully Explained

Muse Code is Meta's terminal coding agent, and Muse Spark 1.2 is the coding-focused model that launched with it on 5 August 2026. Muse Code plans the work, edits code across a whole repository, runs tools and checks its own output. Muse Spark 1.2 costs $1.25 per 1M input tokens and $4.25 per 1M output tokens, or $0.10 and $0.20 on a “contributor” tier where Meta may train on your data. Meta has since released Muse Spark 1.3 (2 September 2026), which replaced 1.2 inside Muse Code at the same price. This guide covers both and shows the numbers that matter.

Muse Code and Muse Spark 1.2 at a glance

• Muse Code: A beta terminal coding agent from Meta Superintelligence Labs, installed with one command.

• Muse Spark 1.2: Released 5 August 2026, a coding-focused update to Muse Spark 1.1 (9 July 2026).

• Context window: The credits are about 1M tokens (1,048,576).

• Standard API price: $1.25 input, $0.15 cached input, $4.25 output per 1M tokens.

• Contributor price: $0.10 input, $0.002 cached input, $0.20 output per 1M tokens, in exchange for letting Meta train on your prompts and completions.

• Weights: Meta's open-weight model is the smaller Muse Glimmer.

• Newer model: Muse Spark 1.3, released 2 September 2026.

Update: Muse Spark 1.3 has replaced 1.2

Four weeks after 1.2, Meta released Muse Spark 1.3 on 2 September 2026. It rolled out in Muse Code and the Meta Model API the same day, and Meta's docs recommend it for new work. 

List prices didn't change: $1.25 and $4.25 for standard, $0.10 and $0.20 for contributor. The model IDs are muse-spark-1.3 and muse-spark-1.3-contributor.

Meta describes 1.3 as stronger on multi-step tool use, browser work, and long-running tasks. One third-party tracker, Artificial Analysis, scored the standard high-effort variant at an Intelligence Index of 61. Treat that and any “fewer tokens per task” claims as early numbers until you've tested them on your own repository.

The benchmark figures below are from Meta's August launch, when 1.2 was the current model. Muse-spark 1.3 scores the same.

What are Muse Code and Muse Spark 1.2?

The language model is Muse Spark 1.2, an extension of 1.1 that has been optimized for coding with additional focus on debugging, understanding the codebase and long developer workflows. Your terminal runs Muse Code as its agent. It is responsible for planning multi-step engineering jobs, editing code across large repositories, running tools, etc., and verifying the result.

They are a pair of structures created by Meta. Muse Spark 1.2 was co-trained in the Muse Code runtime, with agent trajectories and tuning for goal tracking, context compaction and sub-agent coordination, Muse announced. 

What can Muse Code do?

Persistent background agents: Muse Code does not create a new sub-agent for each trivial subtask, but it leaves specialized agents active throughout the session. They identify their own information needs and report back as necessary, so you spend less time on long tasks. Meta claims to have developed six game features concurrently without any collisions, all agents in their own git worktree.

A restart-safe runtime: Muse Code records a local appendable log of all the model calls, tool operations, approvals, and changes to the files. In case of a session crash, it may resume from the very same point. It should matter if the jobs run for a few hours.

Built-in skills: Take a task and make it into a plan that must be approved before any code changes. Run the test plan and the end result keeps the agent running toward the goal through several steps.

Multimodal input: In Meta's demo, Muse Code took a fly-through video of a house as an .mp4 and generated a complete vacation-rental booking page from it.

How to install and start with Muse Code

Meta allows you to install it via this one-liner: curl -fsSL https://dev.meta.ai/install.sh | bash.

Muse Code was launched only for macOS and Linux. Windows users require WSL to be able to launch it. Meta's current pages also include Windows via PowerShell as well, so see dev.meta.ai for the latest platform support.

A sensible first session goes like this. Don't point it at a toy prompt but rather at an actual problem in an actual repository; then, run to get a plan that reflects actual constraints. Execute in advance of approving. Set a scope with a goal for multi-step work. Choose your API tier before any high-volume run, because the two tiers differ by more than ten times in cost.

Muse Spark 1.2 benchmarks

Meta compared Muse Spark 1.2 with Claude Opus 5, GPT-5.6 Terra, Grok 4.5, and Gemini 3.6 Flash, each running in its own preferred agent product at maximum reasoning effort. These are Meta's own numbers.

How to read it: on Terminal-Bench 2.1, Muse Spark 1.2 edges out GPT-5.6 Terra and Grok 4.5 by about a point but trails Claude Opus 5 by almost four. On DeepSWE 1.1, it comes third. On Meta's own internal benchmark, it sits nearly nine points behind Opus 5, which is a candid result for a company presenting its own model.

Each model ran in a different harness, so the scores mix model, harness and prompt effects. Meta says the runs of rival models are best-effort reproductions. 

Muse Spark 1.2 vs Muse Spark 1.1

The gains of about 6.7 and 6.3 points reflect the model, the new harness and the prompts together, not the model alone.

The GPU kernel test

Meta also let Muse Spark 1.2 optimize GPU kernels inside Muse Code over more than 1,000 tool calls and sessions of up to 24 hours, writing, compiling and profiling code on NVIDIA Hopper GPUs. Meta reports steady speedups over the baselines as the run went on. It's a demo of long-horizon and self-correcting work, not a general benchmark.

Muse Spark 1.2 API pricing: standard vs contributor

Both tiers use the same model. The only difference is whether Meta may train future models on your prompts and completions. 

That's roughly 12.5x cheaper on input, 21x on output and 75x on cached input. A job with 100M input and 10M output tokens costs about $167.50 on standard and about $12 on contributor, a saving of more than 92%. A single 40,000-token diff review with 6,000 tokens of feedback costs about $0.076 on standard and about $0.005 on contributor.

Is the contributor tier worth it? It depends on what you send. For proprietary or client code, stay on standard, because the discount is the price you pay in data. The contributor tier also has lower rate limits, which are token-based over a rolling window of about five hours.

The API uses an OpenAI-compatible request format, so many existing integrations only need a new base URL and model name. It accepts text, images, video, audio and PDFs, and returns text only.

Is there a Muse Code subscription?

Besides pay-as-you-go API pricing, Meta's Muse Code page lists subscription tiers aimed at everyday use, with a starter plan shown at around $5 a month and prompt limits per five-hour window. Plan names, prices and limits are the kind of thing that changes quickly, so read Meta's current page before you budget.

Is Muse Spark open source?

No. Muse Spark is not open source and is available only through Meta's products and API. In Meta's release notes, open weights for Muse Spark are a roadmap item. Mark Zuckerberg and Alexandr Wang have said open weights are coming “soon”, but no date has been confirmed.

The open model you can use today is Muse Glimmer, released on 10 August 2026: a 30-billion-parameter model distilled from Muse Spark, under the Apache 2.0 license, small enough to run on a single consumer GPU once quantized.

Muse Code vs Claude Code and Codex

Muse Code arrived about a year after Claude Code and Codex defined this category. It competes with them and Google's Antigravity, xAI's Grok Build and Moonshot's Kimi Code. On Meta's own charts, Claude Opus 5 leads on every coding metric Meta has published, so Muse Code's case is cost, not raw intelligence. If you want to compare the other side, read our guide to Claude Code Review.

Should you use Muse Code?

It's a good fit if you run high-volume, non-sensitive or open-source agentic work where the contributor discount changes your economics; if you want to try persistent parallel sub-agents with worktree isolation; and if you're comfortable with beta software.

Skip it, or at least keep a second tool, if you need the strongest raw coding performance, if you work on client or proprietary code and need the cheap tier, or if you want a battle-tested tool. Muse Code is beta, and its restart-safe runtime hasn't been stress-tested at the scale of the older agents.

Limitations

• Benchmarks aren't apples-to-apples, since each model used its own harness.

• GPT-5.6 Sol and Claude Fable 5 are absent from Meta's comparison.

• It's beta, so expect rough edges and changing platform support.

• Output is text only, even though it accepts images, video, audio and PDFs.

• Contributor-tier rate limits are lower than standard.

• 1.3 now sits above 1.2, so older numbers may not describe what you'll actually run.

  Final verdict

  Muse Code is a promising, early entrant, and the pricing is the real story. Muse Spark 1.2 didn't beat Claude Opus 5 on anything Meta published, but at $0.10 and $0.20 per million tokens on the contributor tier, it changes the math for cost-sensitive, high-volume agent work. With 1.3 now live, start your tests there, keep your data rules clear, and compare the cost of a finished task, not just the rate card.

  Want help choosing coding agents or building AI workflows into your team's process? Our AI Development team builds them, and our AI Consulting team can help you pick the right tools first. Book a free consultation.

  For more on fast-moving AI models, read our Gemini 3.7 Flash guide.

Frequently asked questions

What is Muse Code?

Muse Code is Meta's beta terminal coding agent. It plans, writes and verifies code changes across whole repositories using persistent background agents and a restart-safe execution log.

What is Muse Spark 1.2?

It's Meta Superintelligence Labs' coding-focused update to Muse Spark 1.1, released on 5 August 2026 and co-trained with the Muse Code harness.

How much does Muse Spark 1.2 cost?

Standard pricing is $1.25 per 1M input tokens and $4.25 per 1M output tokens. The contributor tier is $0.10 and $0.20, in exchange for letting Meta train on your prompts and completions.

Is Muse Spark 1.2 free?

There's no free API tier we could confirm. The contributor tier is very cheap but not free. Some third-party tools have offered promotional free access, so check availability before relying on them.

Is Muse Spark 1.2 better than Claude Opus 5?

Not on Meta's own published benchmarks. Opus 5 leads on Terminal-Bench 2.1, DeepSWE 1.1 and Meta's internal coding benchmark. Muse Spark 1.2 does edge out GPT-5.6 Terra and Grok 4.5 on Terminal-Bench 2.1.

What is the difference between Muse Spark 1.2 and 1.3?

1.3 was released on 2 September 2026 and replaced 1.2 in Muse Code. It keeps the same price and context window, and Meta says it improves multi-step tool use and long-running tasks.

Does Muse Code work on Windows?

It launched on macOS and Linux only. Meta's current page lists Windows through PowerShell, so check dev.meta.ai for the latest support.

Are Muse Spark's weights open?

No. Open weights for Muse Spark are a stated roadmap item with no confirmed date. The open-weight model available today is Muse Glimmer (30B, Apache 2.0).

What are the model IDs?

muse-spark-1.2 and muse-spark-1.2-contributor for 1.2, and muse-spark-1.3 and muse-spark-1.3-contributor for 1.3.