Cursor has added Meta’s Muse Wonder AI model to its coding platform, expanding the options available to developers building software with AI agents. The model is designed for complex, multi-step programming work, including debugging, navigating large repositories and executing long sequences of tool calls. Cursor lists Muse Wonder at $4 per million input tokens and $20 per million output tokens, positioning it as a premium option for developers who want an AI model capable of handling extended coding tasks.
The release also puts Meta’s AI research efforts in front of more software developers through Cursor’s existing workflow. Muse Wonder scored 46.1% on CursorBench at its default High reasoning-effort setting, according to Cursor’s published evaluation. The model offers a 300,000-token context window, expandable to one million tokens in Max Mode. However, it is a separately billed third-party model rather than a model whose usage is automatically included in every Cursor subscription.
Key Takeaways
- New model: Cursor has added Meta’s Muse Wonder for agentic coding.
- Pricing: $4 per million input tokens and $20 per million output tokens.
- Benchmark score: 46.1% on CursorBench at the default High effort setting.
- Large context: A 300,000-token context window, with up to one million tokens in Max Mode.
- Coding capabilities: Supports multi-step debugging, file editing, shell commands, browser tools and large-repository work.
- Separate billing: Muse Wonder is a third-party model and may draw on prepaid credits or the applicable usage pool, depending on the Cursor plan.
- Competition: The addition strengthens competition among AI companies seeking to become part of developers’ everyday coding workflows.
What Is Meta’s Muse Wonder?
Muse Wonder is a model developed by Meta Superintelligence Labs, the company’s AI research organisation. It is the successor to Muse Spark 1.3 in Cursor and is designed for work that requires an AI system to maintain context across multiple steps rather than simply generate a short code snippet.
Traditional coding assistants are often used for relatively contained tasks, such as explaining a function, suggesting a fix or completing a few lines of code. Agentic coding models are designed to do more. They can investigate a problem, inspect related files, plan changes, execute commands and use the results to decide what to do next.
Muse Wonder is built around this type of workflow. Cursor says the model can work through long tool-use chains, investigate complex bugs and handle changes across large codebases.
This approach can be useful when a software task does not have an obvious one-line solution. A bug might originate in one component but produce an error elsewhere. Fixing it may require understanding how multiple files interact, running tests and revising the implementation based on the results.
The model’s value will ultimately depend on how reliably it completes those tasks, how much supervision it needs and how much it costs to achieve a usable result.
CursorBench Score: What Does 46.1% Mean?
Cursor reports a score of 46.1% for Muse Wonder at the High reasoning-effort setting on CursorBench 4.0.
CursorBench evaluates AI coding agents on ambiguous, multi-file tasks drawn from real Cursor sessions. The benchmark is intended to measure how well models handle practical software development work involving codebase understanding, investigation, editing, refactoring and other multi-step activities.
A 46.1% score indicates performance on the benchmark’s evaluation tasks; it does not mean the model writes correct code 46.1% of the time in every real-world situation.
| Benchmark detail | Muse Wonder result |
|---|---|
| Evaluation | CursorBench 4.0 |
| Default effort setting | High |
| Reported score | 46.1% |
| Model provider | Meta |
| Primary use case | Agentic coding |
| Context window | 300,000 tokens |
| Maximum context in Max Mode | 1 million tokens |
The score also needs to be considered alongside the model’s reasoning setting. Cursor offers six effort levels: Minimal, Low, Medium, High, Extra High and Max. Higher settings spend more reasoning tokens and can improve benchmark performance, but they may take longer and consume more resources.
Cursor’s published results show that Muse Wonder can reach a 50.4% score at Max effort. That is a different configuration from the default High setting, so the figures should not be treated as directly interchangeable.
Benchmark results are useful for comparing models under a defined evaluation, but they cannot capture every coding environment, programming language, repository or developer preference. A model that performs well on a benchmark may still struggle with a particular project’s architecture or requirements.
How Much Does Muse Wonder Cost?
Cursor lists Muse Wonder at $4 per million input tokens and $20 per million output tokens. Cached input tokens have a separate listed price of $0.20 per million tokens.
Input tokens represent the text and other supported information the model receives, including prompts and relevant code. Output tokens are the content the model generates, such as explanations, proposed edits and code.
| Token category | Price per million tokens |
|---|---|
| Input | $4.00 |
| Cached input | $0.20 |
| Output | $20.00 |
The difference between input and output pricing matters for agentic coding because complex tasks can generate substantial amounts of output. An agent may inspect files, reason through a problem, create edits and produce additional responses across several steps.
For example, a hypothetical workload using one million input tokens and 100,000 output tokens would cost approximately $6 before any applicable plan-specific adjustments: $4 for input and $2 for output. This is an illustration of token pricing, not a prediction of what a typical coding task will cost.
Actual usage can vary considerably. A short code explanation may require relatively little output, while a long debugging session involving repeated edits and tool calls can consume far more tokens.
Cursor also notes that Muse Wonder is a third-party model. Depending on the subscription arrangement, usage is charged against prepaid credits or the applicable pool for other models. Users should check their current plan and billing settings rather than assume that the model is included without additional usage charges.
A Large Context Window for Complex Projects
Muse Wonder supports a 300,000-token context window, with up to one million tokens available in Max Mode at the same per-token rates listed by Cursor.
A context window determines how much information a model can consider within a particular interaction. For coding, that information may include instructions, source files, error logs, test results and documentation.
A larger context can help an agent understand relationships between different parts of a software project. For instance, when a developer changes an authentication function, the agent may need to inspect API routes, configuration files, tests and other components that depend on the same logic.
The ability to consider more information can reduce the need to repeatedly summarise or reload parts of a project. It does not guarantee that a model will correctly understand every file, however. Large contexts still require effective retrieval, prioritisation and reasoning.
Max Mode can be useful for exceptionally large tasks, but developers should not assume that using the maximum context is always necessary. Smaller, well-scoped requests can be faster, easier to review and less expensive to complete.
What Can Muse Wonder Do Inside Cursor?
Muse Wonder can use the tools available to it within Cursor, including searching files and folders, reading code, editing files, running shell commands and controlling a browser.
These capabilities are important because agentic coding requires more than generating text. An agent needs to interact with the development environment to investigate problems and verify whether proposed changes work.
A typical workflow might involve the following steps:
- Investigate: Search the repository and read the files related to a bug.
- Plan: Identify the likely cause and determine which components need changes.
- Edit: Apply the proposed modifications across relevant files.
- Test: Run commands or tests to check the implementation.
- Review: Examine the results and make further changes if needed.
This workflow can help developers delegate repetitive or time-consuming tasks. It can also introduce risks if an agent makes changes without sufficient review. Developers should inspect important edits, run appropriate tests and understand the commands being executed, particularly in production systems or sensitive repositories.
Muse Wonder’s access to browser and shell tools also means that its usefulness depends partly on the permissions and safeguards configured in the development environment. Tool access does not make every action safe or every generated change correct.
How Muse Wonder Fits Meta’s AI Strategy
Meta has been expanding its AI portfolio beyond consumer-facing assistants and into tools that developers can use directly. Making Muse Wonder available through Cursor provides another route for its models to reach people who build software.
Cursor is an important distribution channel because it integrates AI models into an environment developers already use for editing, navigating and testing code. A model that becomes useful in this setting can be evaluated through everyday programming tasks rather than through a standalone chatbot interface alone.
For Meta, adoption by developers could help demonstrate the practical value of its AI research. It also places the company in competition with model providers that offer coding systems through dedicated products, APIs and integrated development environments.
However, availability in Cursor does not establish that Muse Wonder will become a preferred model for every developer. Adoption will depend on performance, reliability, latency, pricing and how well the model fits specific projects.
The competitive landscape also changes quickly. Other models may lead on particular benchmarks, cost less for routine tasks or perform better in certain programming environments. Developers are increasingly able to choose a model based on the job rather than rely on a single provider for every task.
What the Launch Means for AI Coding Tools
The addition of Muse Wonder reflects a wider shift in AI-assisted software development. Coding tools are moving from autocomplete and conversational assistance toward agents that can carry out multi-step tasks with access to files, terminals and browsers.
This creates the possibility of delegating more work to AI, including bug investigation, refactoring, test creation and routine maintenance. But it also changes how developers evaluate productivity. The relevant question is not only how quickly a model produces code, but whether its changes are correct, maintainable and worth the cost of supervision.
Pricing is particularly important for agentic workflows because a model may perform many steps before completing a task. A relatively inexpensive input rate does not necessarily translate into low total cost if the agent generates large amounts of output or repeatedly invokes tools.
Developers should therefore evaluate models using their own workloads. A useful comparison can measure completion rates, time saved, quality of edits, test outcomes and total token spending across a representative set of tasks.
The Bigger Picture
Cursor’s addition of Muse Wonder gives developers another option for agentic coding and broadens Meta’s presence in the software development ecosystem. Its large context window, tool access and 46.1% CursorBench score at High effort make it a model worth evaluating for complex programming tasks. But benchmark results alone do not establish that it will outperform alternatives in every project.
The commercial question is whether the model can complete meaningful work reliably at a cost developers are willing to pay. Separate billing, output-token prices and the amount of supervision required will all affect its practical value. As coding agents become more capable, developers are likely to compare them not only on intelligence but also on the quality, speed and cost of completed work.
Looking Ahead
Developers can access Muse Wonder through Cursor and experiment with different reasoning-effort settings to find the balance between cost and performance that suits their projects. Cursor’s published pricing and benchmark results provide a starting point, but real-world testing will be necessary to determine how the model performs on a particular codebase. Teams should also check provider access and billing settings before using it at scale.
The broader AI coding market will continue to be shaped by competition between model providers and developer-tool platforms. Meta’s latest addition shows that its models are being offered for practical software work, while Cursor’s benchmark and pricing information gives developers a framework for evaluating the option. The long-term winners will need to combine strong coding performance with predictable costs and reliable tool use.
Frequently Asked Questions
1. What is Muse Wonder in Cursor?
Muse Wonder is an AI model from Meta Superintelligence Labs, available in Cursor for agentic coding tasks such as multi-step debugging, repository investigation and code editing.
2. How much does Muse Wonder cost?
Cursor lists the price at $4 per million input tokens and $20 per million output tokens. Cached input tokens cost $0.20 per million.
3. What is Muse Wonder’s CursorBench score?
Muse Wonder scores 46.1% on CursorBench 4.0 at the default High effort setting. Cursor reports a higher score of 50.4% at Max effort.
4. How large is Muse Wonder’s context window?
The standard context window is 300,000 tokens. Max Mode supports up to one million tokens at the same listed per-token rates.
5. Is Muse Wonder included in every Cursor subscription?
No. Cursor classifies it as a third-party model. Billing depends on the applicable subscription and usage arrangement, so users should check their plan details.
6. Is Muse Wonder better than other AI coding models?
The 46.1% score provides a benchmark result at one configuration, not a universal ranking. Performance varies by effort setting, task type and evaluation method. Developers should test models on representative projects before choosing one.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
