Meta Muse Wonder API: How to Use, Pricing, Korea Access

Meta Muse Wonder is an AI model announced by Meta on October 12, 2026. Developers can use it right away through the Meta Model API and Muse Code (Meta's command-line coding tool). The API price is $4 for input and $20 for output per 1 million tokens, the small text units models bill by (as of October 13, 2026). Korea is not listed as a supported or restricted region in Meta's geographic use policy.

What Meta Muse Wonder is

In its official newsroom announcement, Meta described Muse Wonder and Muse Wonder Full as the most capable models it has built so far. They target multi-step, agent-style work for personal and professional tasks. Meta's examples are insurance claims, lease review, and building a working app. They are built on the earlier model, Muse Spark.

The announcement highlights two features. Active Memory carries relevant context across conversations and pulls up the details that fit the current task. Active Focus keeps the goal in view as the work gets harder, tries other approaches, recovers from errors, and checks with the user before important actions. Meta names legal, financial, and health problem interpretation and coding as strengths.

The performance numbers appear as bar charts on the research blog that the announcement points to. Harvey's legal agent benchmark tests whether a model reads source documents and produces legal work product from instructions. SWE-Marathon checks whether a model finishes long-running software projects. FocusBench is an evaluation Meta uses internally, covering multi-step tasks that use tools and files. The scores printed on the charts are below (as of October 13, 2026).

Muse Wonder Full ranks first on all three tests. Muse Wonder leads the comparison models on legal and long-horizon coding, but on FocusBench its 44.9% is lower than GPT-6 Astra (47.7%) and Claude Opus 5.5 (45.8%). Meta ran and published these scores itself, and all comparison models were measured at max reasoning effort (only Muse Wonder Full's SWE-Marathon run used xhigh). Before you adopt it, test your own tasks directly. The announcement says the full safety evaluation report will be published later.

Muse Wonder vs Muse Wonder Full vs Muse Spark

The names are similar, but the routes you can use are different. According to the developer docs, they split as follows.

Model	Where to use it	API model ID	Notes
Muse Wonder	Meta Model API, Muse Code	muse-wonder-1.0	Preview, standard mode only
Muse Wonder Full	Muse Code only	None	Power Usage and Maximum Usage plans
Muse Spark 1.3	Meta Model API, Muse Code	muse-spark-1.3	Default model the docs recommend

According to the Meta Model API model docs, muse-wonder-1.0 and Muse Spark 1.3 both have a context window (the limit on input plus conversation history per request) of 1,048,576 tokens. They accept text, images, video, audio, and PDF as input and output text. Standard mode is the default, in which your input is not used for model training.

How to use the Meta Muse Wonder API

Using the API takes three steps: create a key, make a call, and check the response. In the Meta Model API dashboard, open the API keys menu and click Create API key. Do not write the key into your code. Store it in the environment variable MODEL_API_KEY.

The endpoint is https://api.meta.ai/v1/responses, and you pass the key after Authorization: Bearer in the request header. The API is compatible with the OpenAI SDK, so you can reuse existing code almost as is by changing only the base URL to https://api.meta.ai/v1. The example below is the Python request from the quickstart docs with only the model ID changed to muse-wonder-1.0.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.responses.create(
    model="muse-wonder-1.0",
    input="임대차 계약서의 위험 조항을 5줄로 요약해 줘.",
)
print(response.model_dump_json(indent=2))

The OpenAI SDK does not read the MODEL_API_KEY environment variable automatically, so you must pass api_key yourself as in the code above. If the printed JSON contains a model field and the response text, the call succeeded. A 401 authentication_error means the key is missing or invalid, so check the environment variable value first.

There are two ways to use Muse Code. Pay-as-you-go billing charges per token through your Meta Model API key. A subscription means logging in to the Muse Code CLI with the same account. The subscription docs list four plans: Everyday, High, Power, and Maximum. Everyday gives 10 to 50 prompts every 5 hours, High gives 5 times that, Power 20 times, and Maximum 80 times. Muse Wonder Full is available on Power and Maximum.

Meta Muse Wonder API pricing

Muse Wonder costs 3.2 times more per input token and about 4.7 times more per output token than Muse Spark 1.3 from the same family. The table below shows the dollar rates per 1 million tokens from the official pricing docs. Cached input is the cheaper input rate that applies when you resend the same leading part of a prompt you sent before. Contributor mode lowers the price in exchange for allowing your content to be used for training.

Model (mode)	Input	Cached input	Output
muse-wonder-1.0 (standard)	$4.00	$0.20	$20.00
muse-spark-1.3 (standard)	$1.25	$0.15	$4.25
muse-spark-1.3-contributor (contributor)	$0.10	$0.002	$0.20

There is no surcharge for longer context, and the same rate applies. If you let the model pull evidence through web search, $2.50 per 1,000 searches is charged on top of token costs. Rate limits apply per team, not per key. For Muse Wonder they are 3,000 requests per minute and 4 million tokens per minute.

For example, processing 1 million input tokens and 200,000 output tokens costs $4 plus $4, or $8, on Muse Wonder. On Muse Spark 1.3 it costs $1.25 plus $0.85, or $2.10. This is a simple example of multiplying the rates. Because prices are in dollars, the cost in Korean won depends on the exchange rate at payment time and your card issuer's overseas transaction fee.

Can you use it in Korea

No official document lists Korea as a supported or restricted region. Meta's geographic use policy says only that services are not available in every jurisdiction, and it names one specific restriction: the image model Muse Image is not available in India. What the documents confirm is that Korea is not on the restricted list. Korean language support and domestic Korean payment methods are not covered in any separate guidance.

The experience of using Muse Wonder inside the Muse app has not started yet. This announcement does not mention availability of the app in Korea. The same situation applies to the earlier post on Meta Muse for small businesses: Korea availability.

Muse Wonder or Muse Spark: which to choose

For tasks with simple outputs, such as document summaries, classification, and everyday Q&A, starting with Muse Spark 1.3 is better on cost. Its output rate is about one fifth, and Meta's docs also mark Spark 1.3 as the recommended model. For work where one mistake is expensive to undo, such as legal document review or multi-step coding, send the same input to both models, compare, and then decide whether to move up to Muse Wonder. The model docs mark Muse Wonder as a preview, so check the status label in the model docs again before you lock it into a production service.

The first thing to try is sending the same prompt once to muse-spark-1.3 and once to muse-wonder-1.0, then comparing response quality and token usage side by side. Multiply those token counts by the rates in the pricing table above to get the cost per request. Multiply by your monthly call volume to set a budget.

Common mistakes

• Running the quickstart example as is: the model ID in the docs example is muse-spark-1.3, so you are not using Muse Wonder. Change the model value to muse-wonder-1.0 to be billed at the Wonder rate.
• Assuming max reasoning effort selects Wonder Full: the model docs state that max does not select Wonder Full. Wonder Full is available only to Muse Code subscribers.
• Trying to apply contributor mode to Wonder: Muse Wonder has no contributor mode on pay-as-you-go, so the cheap rate does not apply. The input rate is fixed at $4.
• Budgeting without web search costs: $2.50 per 1,000 searches is added to token costs.

Frequently asked questions

Is content sent to Muse Wonder used for model training

According to the model docs, content in standard mode is not used for training. Contributor mode, which allows training use in exchange for a lower price, exists only for Muse Spark.

When will Muse Wonder be available in the Muse app

The announcement says only that Muse Wonder will soon power the Muse app experience and open to subscribers first. No date is given.