How to Get Started with Meta Muse Wonder: Setup, Pricing, and Traps to Avoid

Meta introduced Muse Wonder as a proprietary frontier model built specifically for extended agent workflows, document analysis, and autonomous tasks. Unlike Meta's open-weight Llama family, Muse Wonder runs behind closed endpoints and subscription tiers, representing a significant shift in Meta's deployment strategy.

Getting started requires understanding the two distinct variants, navigating token economics, and setting up privacy controls properly before sending project code.

Wonder vs. Wonder Full: Choosing the Right Variant

Meta distributes Muse under two separate tiers that serve entirely different user paths.

Standard Muse Wonder (muse-wonder-1.0) is the API model. It handles 1,048,576 tokens of context and accepts multimodal inputs including text, images, audio, video, and PDFs, with text output. Developers can access this version through the Meta Model API and third-party routing aggregators.

Muse Wonder Full is an expanded variant reserved strictly for Muse Code, Meta's coding workspace. Wonder Full cannot be accessed through a pay-as-you-go API. Instead, it requires a flat monthly subscription—either the $50 per month Power Usage plan or the $200 per month Maximum Usage plan. While Wonder Full posts higher scores on complex coding benchmarks, running standard Wonder through the API often delivers plenty of reasoning capacity at a fraction of the cost for routine tasks.

Official Pricing and Token Economics

The standard API tier follows pay-as-you-go billing with aggressive discounts for cached inputs:

Fresh input tokens cost $4.00 per 1 million tokens.
Cached input tokens drop to $0.20 per 1 million tokens, a 95 percent discount.
Output tokens cost $20.00 per 1 million tokens.
Grounded web search queries incur an extra fee of $2.50 per 1,000 requests.

At $4.00 per million input and $20.00 per million output, Muse Wonder sits at a price point comparable to Claude Opus 5.5 and Google Gemini 4 Argon's regular rate. Structuring prompts to reuse persistent context—such as design systems, system instructions, or recurring API schemas—keeps expenses minimal by routing repeated context through prompt caching.

Connecting to the API

The Meta Model API implements OpenAI-compatible routing endpoints at https://api.meta.ai/v1, making integration straightforward without requiring specialized client libraries.

Here is a minimal call using the OpenAI Python SDK (v1.x):

python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=<YOUR_API_KEY>
)

response = client.chat.completions.create(
    model="muse-wonder-1.0",
    messages=[
        {
            "role": "system",
            "content": "You are an agent assistant optimizing code workflows.",
        },
        {
            "role": "user",
            "content": "Explain how prompt caching reduces agent run costs.",
        },
    ],
)

print(response.choices[0].message.content)

When deploying tools and automated commands, provide explicit safety boundaries in your system instructions. While Muse Wonder features Active Memory across long multi-turn sessions, an agent executing bash commands or modifying local files should always require user confirmation before applying destructive changes.

Privacy Controls and Workspace Traps

Account defaults differ depending on whether you access the model via API or through Muse Code subscriptions.

Standard API requests operate under enterprise terms where Meta does not train models on prompts or completions submitted through muse-wonder-1.0. By contrast, Muse Code seats run in Contributor mode by default. In Contributor mode, user code snippets, prompts, and completions may be utilized to train future Meta models.

Teams handling proprietary source code or private customer data should verify workspace settings in Muse Code to disable data sharing before importing repositories, or stick exclusively to the standard API endpoint where data retention exclusions apply automatically.