Meta Muse Wonder: what it is, how much it costs, and how to test the new AI agent

In short: what it is and who needs it

Muse Wonder — a new Meta model for tasks where AI doesn't just respond with a single message, but plans multiple steps, uses tools, modifies files, and verifies results. Meta calls the two main mechanisms Active Memory and Active Focus: the former is meant to carry relevant context between conversations, while the latter continues complex work, tries a different approach after an error, and asks for clarification before important actions.

This could be useful for developers, automation teams, and people building websites or applications with AI. However, feature names are not proof of quality for your project. The practical result of this guide: you will be able to estimate the cost of a single run, prepare a secure task, and compare Muse Wonder with other tools using the same criteria.

What was confirmed by the October 12 release

Meta's official documentation describes Muse Wonder as a multimodal model: it accepts text, images, video, audio, and PDF, and returns text. The stated context window is 1,048,576 tokens. Supported features include tool calling, structured output, search grounding, and reasoning effort levels from minimal to max.

For development, Meta highlights working with large repositories, long agentic processes, and browser or application management. On the model page, the company also mentions planning, delegating subtasks to sub-agents, and error recovery. These are the manufacturer's stated capabilities, not our independent test.

Meta provides results from its own and third-party benchmarks, but a benchmark does not answer the main questions of your project: how many fixes a reviewer will accept, whether an old script will break, how many tokens will be used for retries, and whether the model provides verifiable results. That is why below is a short test on your code.

Muse Wonder, Wonder Full, and API Pricing

At the time of checking, the official price for Muse Wonder in Meta Model API is $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Full is not available via API: the maximum reasoning effort in standard Muse Wonder does not switch the model to Full.

Example arithmetic, not a bill prediction: one run with 100,000 input and 20,000 output tokens would cost about $0.80 — $0.40 for input and $0.40 for output. If all 100,000 input tokens are indeed billed as cached, the math gives about $0.42. The actual amount depends on retries, caching, tools, and billing rules.

Privacy and permissions: what to check before launch

Meta documentation states that in Standard mode, prompts, code, files, context, and responses are not used for training models. For Muse Code subscribers, there is a Contributor mode: it offers more usage in exchange for the ability to train future models on the provided content; the documentation states that this mode is enabled by default and can be disabled in settings. For pay-as-you-go Muse Wonder, Contributor mode is unavailable.

This does not negate basic hygiene. Even when data is not used for training, it may be processed by the service to fulfill the request. Do not share tokens, private keys, personal data, unpublished contracts, or production dumps. Give the agent the minimum necessary permissions: read-before-write, a separate branch, a ban on direct deploy, and manual confirmation for external actions.

30-minute protocol: is the model actually better for you

1. Take one representative task. For example, adding form validation, fixing a mobile menu, or writing a test for an existing function. Do not ask to "improve the whole project."
2. Record the same input. One commit, the same prompt, the list of allowed files, the test command, and the readiness criteria for Muse Wonder and your current tool.
3. Limit the agent's actions. A separate branch, no secrets, no network requests, and no publishing. If an external action is required, let the agent explain it first and wait for confirmation.
4. Record more than just the time. Number of changed files, test results, errors, manual fixes, re-runs, input and output tokens, and final cost.
5. Perform a blind test if possible. A reviewer evaluates two diffs without knowing the model name: correctness, simplicity, security, and adherence to the repository style.
6. Repeat three times. One successful run might be a fluke. For a decision, you need at least three similar tasks, and for production—your usual testing and review process.

Prompt for a controlled agent task

Task template for an AI agentGoal: [one verifiable result]. Context: - repository and branch: [name]; - allowed files: [list]; - forbidden: deploy, changing secrets, external messages, deleting data; - verification command: [exact command]. Readiness criteria: 1. [verifiable behavior]; 2. existing tests pass; 3. a test for the new scenario is added; 4. no new dependencies without explanation. First, read the necessary files and propose a short plan. If data is missing or a dangerous action is required—stop and ask a question. After changes, show the diff, verification results, limitations, and what a human must check manually. Do not claim a test passed if you did not run it.

The template intentionally limits autonomy. A model that can work for a long time can also move in the wrong direction for just as long. Readiness criteria, allowed files, and stop conditions reduce the cost of error.

What not to conclude yet

• "A million tokens" does not mean a million useful tokens. A large context does not guarantee that the model will correctly use every detail of a large repository.
• Active Memory requires a separate privacy assessment. Transferring context between conversations is convenient, but it is important to know exactly what is being stored, how to view it, and how to delete it.
• Long work does not equal correct work. Check intermediate results, tests, costs, and unwanted side effects.
• An additional safety report is still expected. Meta reported that a full report with additional safety evaluations will be published later. The risk picture should not be considered complete without it.
• We have not conducted our own launch of Muse Wonder. This material has verified documentation and arithmetic, but not code quality, availability from a Ukrainian account, or the actual duration of the agent session.

A short checklist before deciding

• Availability and price verified in a personal account.
• Contributor mode is configured according to the team's data policy.
• The test repository contains no secrets or personal data.
• The same task is run on at least two tools.
• Tokens, repetitions, manual edits, and total cost are recorded.
• Tests and manual review completed before merging changes.
• The agent does not have direct production access or uncontrolled external actions.

If Muse Wonder results in fewer errors and manual edits for an acceptable total cost, it is an argument in its favor. If the gain is only in generation speed, but review and correction take more time, the new model has not eliminated the bottleneck.