Last updated: 2026-09-27

Muse Spark API & Muse Code — Meta's Model in Your Own Agent

The Muse app isn't open to developers, but the model behind it is. Meta sells Muse Spark through its own Meta Model API and through OpenRouter. It's a multimodal reasoning model built for long-running agent work, with a 1M-token context window, at a mid-range price. Hermes Agent ships it as a built-in provider, and any agent that speaks the OpenAI or Anthropic API can use it. One tier has a catch worth knowing about.

Models and prices

Model ID	Input / output per M tokens	Cached input	Context	Catch
`muse-spark-1.3` (also 1.2, 1.1)	$1.25 / $4.25	$0.15	1,048,576	None. Meta doesn't train on standard-tier traffic
`muse-spark-1.3-contributor` (also 1.2)	$0.10 / $0.20	$0.002	1,048,576	Meta may train future models on your prompts and completions. 100 requests a minute (standard: 3,000)
`meta/muse-glimmer-30b` (OpenRouter, open weights)	$0.30 / $1.20	$0.04	131,072	A smaller dense model distilled from Muse Spark, Apache 2.0, and runnable locally

Meta Model API and OpenRouter prices checked 2026-09-27. The API also offers Muse Image ($0.01 per image), Muse Voice Transcribe ($0.18 per audio hour) and SAM 3.1. Compare against 40+ models in the cost calculator.

The contributor tier is set by the model ID

It's 92–95% cheaper, and the only difference in your config is the -contributor suffix. Meta's pricing describes it as discounted pricing "in exchange for permission to use your prompts and completions to train future Meta models". If your agent reads email, client documents or code, that data becomes training material. Use it only for public or throwaway data. Hermes asks you to confirm before selecting a contributor model interactively; other tools may not warn you.

Direct from Meta: the Meta Model API

Base URL https://api.meta.ai/v1, with a bearer key from dev.meta.ai. It offers OpenAI-compatible Chat Completions and Responses endpoints and an Anthropic-compatible Messages endpoint, all with the same key and billing:

curl https://api.meta.ai/v1/chat/completions \
  -H "Authorization: Bearer $MODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"muse-spark-1.3","messages":[{"role":"user","content":"Hello"}]}'

In Hermes Agent (built in since v0.21.0)

# ~/.hermes/.env
MODEL_API_KEY=...

hermes chat --provider meta-ai --model muse-spark-1.3
# or choose Meta Model API in: hermes model

Hermes routes api.meta.ai through the Responses API so prompt caching works. See the Hermes model guide.

In OpenClaw or NemoClaw (via OpenRouter)

OpenRouter serves meta/muse-spark-1.3 at the same price as Meta. In OpenClaw, add an OpenRouter key and set the model to Muse Spark through the OpenRouter provider. OpenClaw validates its config strictly, so run openclaw config validate after editing (see configuration). In NemoClaw, switch the route on the host; the key stays outside the sandbox:

nemoclaw inference set --provider openrouter-api --model meta/muse-spark-1.3 --sandbox my-assistant

Details: switching providers in NemoClaw. Test tool calling on your own workload before switching an always-on agent.

Muse Code: Meta's coding agent

Muse Code is Meta's terminal coding agent for Muse Spark (early beta, announced 2026-08-05 with Muse Spark 1.2). It's built for long, multi-agent coding runs, and every subagent, tool call, steer and cancel can be replayed from an event log.

curl -fsSL https://dev.meta.ai/install.sh | bash
cd your-project && muse

It has its own plans: Everyday (10–50 prompts per 5 hours), High (5× that) and Power (20×). They're separate from Muse app subscriptions and require you to be 18+. Or pay per token through the API. Hermes borrowed one idea from it: optional git-worktree isolation for subagents.

Is Muse Spark a good agent model?

Strengths: 1M context at $4.25 per million output tokens, fast, multimodal input (text, images, video, audio, PDF), and built for long agent runs.
Limits: reviewers put it below the top frontier models (GPT-6 Astra, Claude Fable 5.1, Opus 5.5) on the hardest reasoning. See our benchmarks.
Where it fits: a mid-price default for long-context agent work, or a background-job model priced between Claude Haiku 4.5 ($1/$5) and Sonnet 5 ($2/$10) on output.

More Meta Muse guides

Hub · Getting started · Pricing & limits · Privacy & security · Muse Spark API & Muse Code · Muse vs Grok Bot, Hermes & ChatGPT

Sources: Meta Model API overview, pricing and rate limits, Muse Spark 1.2 and Muse Code announcement (2026-08-05), OpenRouter's model catalogue, and Hermes Agent's provider docs. Checked 2026-09-27.