How to Use Muse Wonder: OpenRouter, Meta API, and Muse Code

👁 119 reads

Muse Wonder has been available since launch day, but not necessarily where most people look for it. It is not powering Meta AI inside WhatsApp, Instagram, or Messenger yet, and the standalone Muse app remains restricted outside the US and Canada. You can, however, run it through OpenRouter or Meta’s official API, while the more capable Muse Wonder Full variant is available inside the Muse Code terminal agent. Here is a step-by-step walkthrough covering all three paths, along with a working Python script tested on October 12, 2026. For a complete model breakdown, see our Muse Wonder review.

Comparing Access Methods

Method 1: OpenRouter Web Interface Without Coding

OpenRouter operates as an aggregator offering a single API key and chat interface across hundreds of models. Here is how to set it up:

1. Go to openrouter.ai and sign up using Google, GitHub, or email.
2. Under the Credits tab, add funds using a payment card. A few dollars will comfortably cover dozens of test runs.
3. Open Chat, open the model selector, and search for “Muse Wonder”. Select “Meta: Muse Wonder”.
4. If prompted for 18+ age verification, complete it in your account settings (OpenRouter previously enforced this for Meta models during the Muse Spark launch).
5. Start chatting just like ChatGPT. You can attach PDFs, images, or audio files - the architecture natively processes text, image, video, and audio inputs.

Keep inference costs in mind: Muse Wonder always outputs reasoning tokens, even at lower effort settings. A single-sentence prompt asking to calculate “17% of 2350” consumed 253 output tokens ($0.005). Across our 13-task test suite, internal reasoning accounted for approximately three-quarters of all generated output tokens.

Method 2: OpenRouter API - Ready-to-Use Python Script

The Python script below sends a prompt to Muse Wonder and outputs both the generated response and the exact execution cost. Tested on October 12, 2026 using Python 3.13 on Windows 11, the script accurately handled our test prompt with an execution cost of $0.005156.

1. Install Python 3.10+ and the requests library: pip install requests.
2. Generate an API key in OpenRouter (Keys tab) and assign it to an environment variable. On Windows PowerShell: $env:OPENROUTER_API_KEY="your-api-key"; on macOS/Linux: export OPENROUTER_API_KEY="your-api-key".
3. Save the code snippet below as muse_wonder.py.
4. Execute it via terminal: python muse_wonder.py "Your question here".

# Muse Wonder via OpenRouter - minimal script (tested Oct 12, 2026)
# Requirements: pip install requests; environment variable OPENROUTER_API_KEY
import os, sys, requests

# Note: The original test used a Polish fallback prompt:
# "Napisz po polsku krotki mail z przeprosinami za opozniona dostawe."
# Translated below for reference:
fallback_prompt = "Write a short email apologizing for a delayed delivery."

prompt = " ".join(sys.argv[1:]) or fallback_prompt
r = requests.post(
    "https://openrouter.ai/api/v1/chat/completions",
    headers={"Authorization": "Bearer " + os.environ["OPENROUTER_API_KEY"]},
    json={
        "model": "meta/muse-wonder",
        "reasoning": {"effort": "low"},   # minimal, low, medium, high, xhigh, max
        "usage": {"include": True},
        "messages": [{"role": "user", "content": prompt}],
    },
    timeout=300,
)
r.raise_for_status()
d = r.json()
print(d["choices"][0]["message"]["content"])
u = d.get("usage", {})
print(f"\n[tokens: prompt {u.get('prompt_tokens')}, completion {u.get('completion_tokens')}, cost {u.get('cost')} USD]")

You can adjust the depth of thinking via the effort parameter. Using low returns faster, cheaper responses, whereas max instructs the model to exhaustively reason through tasks - the setting Meta utilized when recording its official benchmark scores.

Method 3: Meta Model API

Directly on Meta’s platform, the model identifier is muse-wonder-1.0. Pricing matches OpenRouter: $4 per million prompt tokens, $20 per million completion tokens, and $0.20 per million cached prompt tokens. Rate limits sit at 3,000 requests and 4 million tokens per minute per team. Developer accounts are registered via dev.meta.ai. Note that Meta has not published an exhaustive list of country availability for this endpoint, and certain European users experienced regional geo-blocking warnings earlier in the year. Meta has also not published an EU data processing addendum (DPA) for this service.

Method 4: Muse Code and Wonder Full

The flagship variant, Muse Wonder Full (scoring 56 points on the Artificial Analysis index), is exclusively accessible through Muse Code, Meta’s terminal-based coding agent. Subscription plans:

• Everyday Usage: $5 per month;
• High Usage: $15 per month;
• Power Usage: $50 per month, includes Wonder Full;
• Maximum Usage: $200 per month, includes Wonder Full.

Muse Code can also be connected via API token billing, but doing so routes queries through standard Wonder rather than the Full release.

Current Limitations and Restrictions

• Meta AI in WhatsApp, Messenger, and Instagram: Does not feature Muse Wonder anywhere yet, including the US. Meta plans to roll the model into these consumer surfaces gradually, prioritizing paying Muse subscribers first.
• Muse App: The full personal agent (handling inbox and calendar scheduling for $20 to $100 per month) remains restricted to users in the United States and Canada. Accessing it from other territories produces a regional restriction error.
• Local Model Weights: Muse Wonder is closed-source. Meta’s open weights ecosystem remains anchored by Llama and Muse Glimmer.

Frequently Asked Questions

Can I use Muse Wonder for free?

No. You must pay per token ($4 / $20 per 1M tokens) via an API provider or hold an active paid Muse Code subscription. There is currently no free tier.

Do I need a VPN to access it?

Not for OpenRouter, which accepts international accounts directly. A VPN will generally not unlock the consumer Muse app either, as Meta validates local payment methods and phone numbers in addition to IP location.

How does Muse Wonder perform in other languages?

Multilingual capabilities are strong. In our 13-task benchmark conducted entirely in Polish, the model scored 87.0 points, securing 2nd place in our historical rankings.

// read moreMuse Wonder Pricing: API, Muse Code, and Real Test CostsAI ToolsMeta Muse Wonder: Benchmarks, API Pricing, and Full TestBenchmarksOpenRouter Guide: One API Key for 460+ AI ModelsGuides