Guide
OpenClaw + Muse Spark 1.2: Meta's 1M-Context Agent Model
Meta shipped Muse Spark 1.2 on 5 August 2026 — a reasoning model built for agentic work, with a 1M-token context window and pricing that undercuts most models in its class. It is on OpenRouter today, which means an OpenClaw agent can use it with one key. Here is the model ID, the honest pricing, and the one setting you should think about before you save it.
What Muse Spark 1.2 Is
Muse Spark 1.2 is a reasoning model from Meta Superintelligence Labs, released alongside Muse Code, Meta's terminal coding agent. It accepts text, images, video, audio, and PDF documents and returns text, with a 1M-token context window.
The weights are not published. It is a proprietary model reachable only through Meta's own API, its apps, or partner platforms — so unlike an open-weights release, there is no self-hosted route. For an OpenClaw agent, the practical path is OpenRouter.
Specs and Pricing
| Property | Value |
|---|---|
| Model ID (OpenRouter) | meta/muse-spark-1.2 |
| Context window | 1M tokens |
| Input modalities | Text, image, video, audio, PDF → text out |
| Standard price | $1.25 / 1M input, $4.25 / 1M output |
| Cached input | $0.15 / 1M |
| Contributor tier | $0.10 / 1M input, $0.20 / 1M output — see the trade-off below |
| Weights | Proprietary, not published |
At $1.25 in and $4.25 out with a million-token window, Muse Spark 1.2 sits in an unusual spot: frontier-adjacent capability at roughly a third of what the priciest flagships charge on input. That is the reason it showed up in agent circles within days of launch.
The Contributor Tier Is Not Free Money
Meta offers muse-spark-1.2-contributor at $0.10 input and $0.20 output — roughly 12× cheaper on input and 21× cheaper on output. The price cut is not a promotion. You are granting Meta permission to train future models on your prompts and completions.
Benchmarks
Meta published Muse Spark 1.2 scoring 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1. On Meta's own comparison table both figures are second place, behind Claude Opus 5.
Read that the way it is meant: these are vendor-published numbers, and a vendor chart that puts its own new model second is at least candid about where it lands. Terminal-Bench measures agentic CLI work, which is the closest public proxy for what an OpenClaw agent actually does all day — so it is the more relevant of the two here.
How to Use Muse Spark 1.2 on OpenClaw Launch
Muse Spark is new enough that it is not yet a curated one-click pick. Check the model list for what is available as a single click today. Until it lands there, run it as a bring-your-own-key model through OpenRouter:
- Create an OpenRouter key at openrouter.ai/keys. No minimum spend — you pay per token.
- Add the key on your API Keys page, or paste it in the configurator under BYOK. See the BYOK guide for the full walkthrough.
- Set the model to
meta/muse-spark-1.2. - Pick your channel (Telegram, Discord, WhatsApp, WeChat, or the web gateway) and click Deploy.
You can switch to it at runtime from any connected chat with /model meta/muse-spark-1.2 — no redeploy needed.
Self-Hosted Configuration
Running your own OpenClaw? OpenRouter is already a supported provider, so this is a model-ID change in openclaw.json:
{
"models": {
"providers": {
"openrouter": {
"apiKey": "sk-or-..."
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "openrouter/meta/muse-spark-1.2"
}
}
}
}Full provider setup is covered in the OpenClaw + OpenRouter guide.
A Note on Availability
OpenRouter is Meta's partner for Muse Spark, and requests are forwarded to a single upstream provider rather than load-balanced across several. That matters for one reason: when the upstream has a bad day, there is no second provider to fall back to. If your agent must answer around the clock, keep a second model configured — see best models for OpenClaw for sensible pairings.
Muse Spark 1.1 launched limited to US developers. 1.2 shipped with expanded global access to both Muse Spark models, so the earlier regional restriction no longer applies the way it did in July.
FAQ
Can I self-host Muse Spark 1.2?
No. The weights are not published. It is available only through Meta's API, apps, or partner platforms such as OpenRouter. If self-hosting is the requirement, look at an open-weights model instead — see OpenClaw + Ollama.
Is the contributor tier safe for a work bot?
Not by default. The discount is granted in exchange for permission to train on your prompts and completions, and a shared work bot reads other people's content. Use the standard tier for anything handling data that is not exclusively yours.
Do I need a Meta account?
Not for this route. Through OpenRouter you use an OpenRouter key and pay OpenRouter. A Meta developer account is only needed if you call Meta's API directly.