Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, but the two models are not available in the same way. Muse Wonder is available to developers through the Meta Model API and through Muse Code, Meta’s coding agent for the terminal. Muse Wonder Full is available only inside Muse Code, and only on eligible subscription plans. This guide explains how to use Meta Muse Wonder step by step, including installation, API setup, a first request, prompting, pricing, and troubleshooting.
Quick Answer: How Do You Use Meta Muse Wonder?
You can use Muse Wonder in two ways. The first is Muse Code: install the command-line tool, sign in, and start it with the model set to muse-wonder-1.0. The second is the Meta Model API: create an API key, then send requests to https://api.meta.ai/v1 with the same model ID. Muse Wonder Full is not available through the API. It is included only with eligible Muse Code subscriptions.
Key takeaways:
- Use Muse Code to work on a software project from your terminal, or use the Meta Model API to add the model to your own application.
- The Muse Wonder model ID is muse-wonder-1.0.
- Muse Wonder Full requires a Muse Code Power Usage or Maximum Usage plan.
- Setting reasoning effort to the maximum on Muse Wonder does not switch you to Muse Wonder Full.
- API billing is per token: $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens.
- Review and test every generated change before you rely on it.
What Are Meta Muse Wonder and Muse Wonder Full?
Meta built both models for agentic work, where a model plans several steps, uses tools, checks its progress, and recovers when something fails. Meta highlights complex coding, multi-file changes, long-running agents, tool use, computer use, and professional tasks such as reviewing documents. Muse Wonder Full is the higher-capability variant of the family.
Two features support this work. Active Memory carries relevant context across conversations and brings back the details a current task needs. Active Focus helps the model keep working toward a goal when a task gets difficult. According to Meta, it tries other approaches, works through blockers, and asks for clarification or checks in before important actions.
These features matter most for tasks with many steps. Answering one question needs one good response. Adding a feature across several files requires the model to track the goal, remember earlier decisions, run tests, and fix what breaks.
Meta’s benchmark results are its own reported measurements, and Meta cites some coding results from third-party runs. Treat all reported results as vendor claims. Muse Wonder is in preview, so test it on your own work before depending on it.
Meta Muse Wonder vs. Wonder Full: Key Differences
| Feature | Muse Wonder | Muse Wonder Full |
| Availability | Meta Model API and Muse Code | Muse Code only, with an eligible subscription |
| Access method | API calls to https://api.meta.ai/v1, or Muse Code on pay-as-you-go or any plan | Muse Code Power Usage or Maximum Usage plan |
| Model ID | muse-wonder-1.0 | No API model ID. Not available through the API |
| Context window | 1,048,576 tokens | 1M context, as listed on Meta’s model page |
| Pricing | API: $4.00 input, $0.20 cached input, $20.00 output per million tokens. Muse Code: pay-as-you-go or subscription | Included in a subscription. Power Usage is $50/month; Maximum Usage is $200/month |
| Best-fit workloads | API integrations, multi-file coding changes, long agent runs, sub-agent orchestration | The most demanding coding and agent workflows inside Muse Code |
| Important limitations | Preview status. Standard mode only on the API, with no Contributor mode | Subscription only. Usage limits apply. Cannot be called from your own code |
Setting reasoning effort to max on Muse Wonder does not select Muse Wonder Full. The two are separate variants.
Recommendations:
- For API integration, choose Muse Wonder. It is the only version you can call from code.
- For demanding coding inside Muse Code, choose Muse Wonder Full if your subscription includes it. Otherwise, Muse Wonder works on pay-as-you-go or any Muse Code plan.
- For cost-sensitive projects, estimate your token use first. Meta lists Muse Spark 1.3 at $1.25 per million input tokens and $4.25 per million output tokens in Standard mode, and describes Muse Spark as the lower-cost option for high-volume work.
How to Access Meta Muse Wonder
There are two routes. Choose Muse Code if you want an agent that works inside a software project from the terminal. Choose the API if you are writing software that calls the model, such as an internal tool, an automation, or a product feature.
Option 1: Use Muse Code
Muse Code is a command-line coding agent. It plans tasks, edits files, and runs commands, with approval prompts and a sandbox turned on by default. It runs on macOS, Linux, and Windows. You can sign in with a browser or an API key, and you can pay per token or with a monthly subscription.
Option 2: Use the Meta Model API
Create an account at dev.meta.ai, add a payment method on the Billing page, and create an API key in the dashboard. Billing is usage-based, with no minimums or upfront commitment. Before writing code, read the Muse Wonder model page and the getting-started guide.
How to Use Meta Muse Wonder Through Muse Code
Step 1: Install Muse Code
On macOS or Linux, open Terminal and run:
bash
curl -fsSL https://dev.meta.ai/install.sh | sh
On Windows, open PowerShell and run:
powershell
irm https://dev.meta.ai/install.ps1 | iex
Confirm the installation with muse –version. If the command is not found, open a new terminal window and try again.
Step 2: Sign In and Check Your Plan
Change into your project folder and run muse. On first run, choose a browser sign-in or paste an API key. A browser sign-in is the simpler option for individual use.
To check or change your subscription, sign in at dev.meta.ai, then open Subscriptions in Accounts Center and select Muse Code. Muse Code subscriptions work only through the Muse Code CLI.
Step 3: Select the Model
Start Muse Code with the model ID:
bash
muse –model muse-wonder-1.0
You can switch models during a session with /models. Reasoning effort defaults to high. For the hardest problems, add –reasoning-effort max, which produces slower responses and uses more of your usage allowance.
The documentation I reviewed does not publish a launch flag for Muse Wonder Full. If your plan includes it, open the model list with /models in an interactive session, confirm the exact name, and then start the task.
Step 4: Give the Model a Defined Task
A clear prompt includes a goal, constraints, and acceptance criteria. The following example assumes a small Python command-line todo app:
Text
Goal: Add a “done” command to this todo CLI. “todo done <id>” marks a task complete. “todo list –open” hides completed tasks.
Constraints: Keep the storage format in data/tasks.json unchanged. Keep existing command names. Use only the standard library.
Acceptance criteria: All existing tests pass. Add tests for both commands. Your final message lists every file you changed.
Process: Before editing, give me a short plan and wait for my approval.
Step 5: Review, Run, and Test the Result
Muse Code does not commit, amend, or push changes unless you ask it to. If the project uses git, run git diff to inspect every change. Run your full test suite, and check any new dependencies the agent added.
Read approval prompts before you accept them. If a test fails, reply with the exact error output and the behavior you expected. Generated code can look correct and still fail on edge cases, so the tests are your evidence.
How to Set Up and Use the Muse Wonder API
Meta states that the Model API is drop-in compatible with the OpenAI SDK. The steps below use the OpenAI Python package.
Step 1: Generate an API Key
Sign in at dev.meta.ai, open the API keys tab, click Create API key, and copy the key. You can view it only once. Create a separate key for each application, and never commit keys to a repository.
Step 2: Prepare Your Environment
You need Python 3.9 or later. Install the SDK:
bash
pip install openai
Store your key in an environment variable. On macOS or Linux:
bash
export MODEL_API_KEY=”your-api-key”
In PowerShell on Windows:
powershell
$env:MODEL_API_KEY = “your-api-key”
Meta’s quickstart notes that the OpenAI SDK does not read MODEL_API_KEY automatically, while the authentication page says official SDKs do. The example below passes the key explicitly, which works either way.
Step 3: Configure the Client
Use these values:
- Base URL: https://api.meta.ai/v1
- Model ID: muse-wonder-1.0
For agentic work, Meta recommends the Responses API, which carries reasoning across turns.
Step 4: Run a First Request
python
import os
from openai import APIError, OpenAI
client = OpenAI(
base_url=”https://api.meta.ai/v1″,
api_key=os.environ[“MODEL_API_KEY”],
)
prompt = (
“Write a Python function slugify(title: str) -> str. “
“Lowercase the input, replace each run of non-alphanumeric characters “
“with one hyphen, and strip leading and trailing hyphens. “
“Return only the code with a short docstring.”
)
try:
response = client.responses.create(
model=”muse-wonder-1.0″,
reasoning={“effort”: “medium”},
input=prompt,
max_output_tokens=4000,
)
print(response.output_text)
except APIError as err:
print(f”Request failed: {err}”)
Reasoning tokens count toward max_output_tokens, so set the limit high enough to cover both reasoning and the visible answer. A successful call prints the model’s text. A 401 authentication_error means the key is missing or invalid. A 404 model_not_found means the model ID is misspelled.
Step 5: Extend the Example Into a Workflow
For multi-turn conversations, pass the previous response’s id as previous_response_id. Tool calling requires your code to define each tool and execute it. Search grounding is available by adding a web_search tool, billed at $2.50 per 1,000 queries on top of token costs. Meta’s cookbook includes recipes for computer use, sub-agent fan-out, and GitHub agents. Each one requires you to build the surrounding integration and set permissions.
Practical Ways to Use Muse Wonder and Wonder Full
The scenarios below are illustrative. They are not results from tested experiments.
1. Build and Test a Small Application
Problem: Turning an idea into a working prototype with tests. Example prompt: “Build a command-line expense tracker in Python with add, list, and total commands. Include tests for each command.” Verification: Run the tests and try each command with sample data. Route: Muse Code. Limitation: Generated apps need review for input validation and security.
2. Debug a Multi-File Codebase
Problem: A failing test whose cause spans several modules. Example prompt: “test_invoice_total fails. Find the root cause, make the smallest fix, and list the files you changed.” Verification: Run the full suite and confirm the original failure no longer occurs. Route: Muse Code. Limitation: Tests only catch bugs they cover, so the fix can still miss related issues.
3. Refactor Legacy Code
Problem: Code that works but is hard to change. Example prompt: “Refactor the report module into smaller functions without changing its output. Add a golden-output test before you change anything.” Verification: Compare outputs before and after the refactor. Route: Muse Code. Limitation: Split large refactors into reviewable stages instead of accepting one large diff.
4. Automate a Multi-Step Research or Business Workflow
Problem: A repetitive process, such as reading a CSV, summarizing rows, and drafting a report. Example approach: A script that reads the file, sends each batch to muse-wonder-1.0 with fixed instructions, and writes the summaries to disk. Verification: Check a sample of summaries against the source data. Route: API. Limitation: Model output can contain errors, so people should review figures before they are used. Large inputs increase cost.
5. Build a Tool-Using Agent With Limited Permissions
Problem: An assistant that takes actions, such as creating tickets. Example approach: Define each tool with a narrow scope, and require human confirmation before any write action. Verification: Test each tool against a sandbox account and log every tool call. Route: API with tool calling. Limitation: The model can access only what your code grants it.
How to Write Better Prompts for Muse Wonder
Effective prompts for these models usually do six things: state the objective and output, provide relevant context and constraints, define acceptance criteria, request a plan for complex work, require tests or evidence, and set approval boundaries for risky actions. Use this template:
Text
Goal: [the outcome in one sentence]
Context: [relevant files, versions, or background]
Constraints: [what must not change, required libraries, style rules]
Acceptance criteria: [tests to pass, behavior to verify]
Process: Give a short plan first. Pause before deleting files, changing dependencies, or running commands that affect data. End with a list of changed files and the test results.
The template works because explicit acceptance criteria define a finish line, a requested plan exposes misunderstandings early, and the approval boundary matches Muse Code’s approval system.
How Much Does Muse Wonder Cost?
| Option | Price | Notes |
| Muse Wonder API input | $4.00 per 1M tokens | Standard mode |
| Muse Wonder API cached input | $0.20 per 1M tokens | Applies to tokens matched to a cached prefix |
| Muse Wonder API output | $20.00 per 1M tokens | Reasoning tokens are billed as output |
| Web search grounding | $2.50 per 1,000 queries | Added to token costs |
| Muse Code Power Usage | $50/month | Includes Muse Wonder Full |
| Muse Code Maximum Usage | $200/month | 80x the Everyday Usage limit |
Your API cost depends on input, cached input, and output tokens. For example, a hypothetical request with 20,000 input tokens (15,000 of them cached) and 2,000 output tokens would cost about $0.063: $0.02 for 5,000 uncached input tokens, $0.003 for cached input, and $0.04 for output.
Subscription plans use a flat monthly rate with usage limits, which is a different billing model from API token rates. Meta’s model page lists prices for Power Usage and Maximum Usage. Check Accounts Center for current prices of the other plans. Muse Code also supports pay-as-you-go billing through your API key.
Meta’s pricing page also lists Muse Spark rates, which are lower than Muse Wonder’s. Compare them before you choose a model for high-volume work.
Common Muse Wonder Problems and How to Fix Them
- Authentication failures (401): Confirm that MODEL_API_KEY is set in the same shell where you run your script. If a key was deleted or lost, create a new one, since keys appear only once.
- Wrong model ID or unavailable model (404): Check that the model ID is exactly muse-wonder-1.0. Send a request to GET https://api.meta.ai/v1/models to see which models your account can use.
- Plan does not include Muse Wonder Full: Confirm that your subscription is Power Usage or Maximum Usage and that you are using Muse Code. Muse Wonder Full cannot be called through the API.
- Invalid requests or SDK compatibility errors: Compare your parameters with the Responses API reference. Some OpenAI parameters, including stop, logit_bias, and n values above 1, are not implemented on this endpoint.
- Unexpected token costs: Check the usage object in each response, count tokens before sending large inputs, and check the usage dashboard at dev.meta.ai/usage. Lower the reasoning effort for simple tasks.
- Incomplete output: Raise max_output_tokens. Reasoning and visible output share that budget.
- Agent workflows fail: Check the approval mode and sandbox settings. By default, the sandbox limits filesystem writes and allows network access only after approval for each new destination.
Safety, Privacy, and Limitations
Generated code needs human review and testing. Muse Code asks before risky actions and runs shell commands in a sandbox by default. The –yolo and –disable-sandbox flags remove those protections, so use them only in isolated environments.
Data use depends on the mode. In Standard mode, Meta does not use your prompts, code, files, or responses to train its models. Muse Code subscriptions default to Contributor mode, which gives more usage in exchange for permission to use your content for training. You can turn it off in Muse Code settings. Contributor mode is not available for Muse Wonder on pay-as-you-go API billing. Verify your current settings before you work with sensitive material.
Avoid submitting confidential business or personal information until you have reviewed Meta’s terms and policies for your account. Benchmark scores do not guarantee performance on your tasks. Preview status, usage limits, country availability, and features may change, so confirm current details before you rely on them.
Final Verdict
Choose Muse Wonder through the API when your application needs to call the model programmatically. Choose Muse Code when you want an agent that works inside a software project from the terminal. Choose Muse Wonder Full when your workflow requires it and your eligible subscription includes it. In every case, start with a small task, verify the output, watch your usage, and expand only after the results hold up.
Frequently Asked Questions
How do I use Meta Muse Wonder?
Install Muse Code and run muse –model muse-wonder-1.0 in a project folder. To call the model from code, create an API key at dev.meta.ai and send requests to https://api.meta.ai/v1 with the model ID muse-wonder-1.0.
How do I access Muse Wonder Full?
Subscribe to a Muse Code Power Usage or Maximum Usage plan through Accounts Center, then use it in the Muse Code CLI. Muse Wonder Full is not available through the Meta Model API.
Is Meta Muse Wonder free to use?
Meta’s documentation does not describe a free tier for Muse Wonder. The API bills per token at published rates, and Muse Code runs on pay-as-you-go billing or a monthly subscription. Confirm current pricing before you start.
Can I use Muse Wonder without coding?
Muse Code is a terminal tool designed for software projects, so it suits people who are comfortable running basic commands. Meta says Muse Wonder will soon power experiences in Muse, rolling out first to subscribers.
Can I access Muse Wonder Full through the API?
No. Meta’s model page lists Muse Wonder Full as unavailable through the Model API. It is available only in Muse Code on eligible subscription plans, and raising reasoning effort on Muse Wonder does not change that.
What is the difference between Muse Wonder and Wonder Full?
Muse Wonder is available through the API and Muse Code. Muse Wonder Full is available only in Muse Code on Power Usage and Maximum Usage plans. Meta describes Wonder Full as its most advanced variant, though its comparisons are vendor-reported.
What is the Muse Wonder API model ID?
The model ID is muse-wonder-1.0, and the base URL is https://api.meta.ai/v1. Muse Wonder is in preview, so check the model list before deploying to production.
How much does the Muse Wonder API cost?
Standard-mode Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. Web search grounding adds $2.50 per 1,000 queries. Total cost depends on your token usage, so measure it before you scale.