How to Use Muse Wonder Across APIs, IDEs, and Apps

On October 12, 2026, Meta introduced the Muse Wonder series, a family of frontier artificial intelligence models built for personal and professional agent tasks. Learning how to use Muse Wonder starts with understanding the two variants: the base Muse Wonder model and the larger Muse Wonder Full model. Both systems are designed to plan multi-step workflows, call tools, write multi-file code, and operate desktop and mobile software.

Unlike standard text interfaces like ChatGPT or Claude that focus primarily on conversational turns, Muse Wonder incorporates active focus and active memory across long tasks. Meta built these models on larger base architectures trained at its Prometheus data center to succeed its earlier Muse Spark line. The models can run autonomous workflows for hours, hand off tasks to sub-agents, and maintain context across sessions using a 1 million token context window.

For operators in compliance-heavy fields like law, accounting, and healthcare, this release changes how teams evaluate automated research and document intake. Meta evaluated the architecture on specialized legal, financial, and clinical benchmarks like Harvey LAB-AA and Vals Finance Agent V2. Understanding where each model variant runs helps technical teams adopt agentic automation without creating governance risks.

How to Use Muse Wonder Across Available Surfaces

You can access Muse Wonder through three distinct environments depending on your technical requirements and subscription tier. Meta distributes the standard model, identified as muse-wonder-1.0, through the Meta Model Application Programming Interface (API) and third-party aggregators such as OpenRouter. Developers writing software can also interact with it directly inside Muse Code, Meta's dedicated development environment.

Access to the larger Muse Wonder Full model, identified as muse-wonder-full-1.0, requires a paid Muse Code subscription on the Power Usage or Maximum Usage tier. Meta does not offer Muse Wonder Full through a standalone pay-as-you-go API endpoint at this time. Teams that require the larger model must route their workloads through the Muse Code developer interface rather than custom backend scripts.

Consumer web and mobile access will follow a staggered deployment schedule. Meta announced that Muse Wonder will soon power experiences in Muse, its consumer personal agent application, beginning with existing paid subscribers. If you only use consumer browser interfaces, you must wait for Meta to activate the model on your account.

• Meta Model API: Provides programmatic access to the standard muse-wonder-1.0 model using endpoints compatible with the OpenAI Software Development Kit (SDK).
• OpenRouter: Provides third-party routing and token management for the muse-wonder-1.0 model without a direct Meta developer agreement.
• Muse Code IDE: Supports both muse-wonder-1.0 and muse-wonder-full-1.0 for multi-file repository maintenance and background terminal commands.
• Muse Consumer App: Staged rollout scheduled for web and mobile agent tasks, prioritizing existing subscribers before free users.

Pricing Plans and Subscription Requirements

Running Muse Wonder via the Meta Model API costs $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. Prompt caching reduces the cost of large system instructions and reference documents by 95 percent when requests share context. Teams processing recurring legal filings or complex policy handbooks should structure requests to keep cache prefixes identical.

Running Muse Wonder Full does not use per-token billing. Meta includes access to the full model exclusively within Muse Code subscriptions, which cost $50.00 per month for the Power Usage plan and $200.00 per month for the Maximum Usage plan. The Power Usage plan provides 20 times the usage allowance of the base Everyday Usage tier, while the Maximum Usage plan provides 80 times the base allowance along with priority task queues.

These developer subscriptions remain distinct from the consumer Muse personal agent plans. Meta offers consumer Muse tiers at $20.00 and $100.00 per month, which do not include API keys or Muse Code integration privileges. Make sure your accounting team separates API infrastructure billing from individual desktop seat licenses to avoid unexpected credit limits.

How to Use Muse Wonder with the Meta Model API

Configuring the Meta Model API requires passing the model identifier muse-wonder-1.0 into an environment compatible with the standard OpenAI SDK format. Because Meta adopted standard endpoint conventions, you can swap existing API clients by changing the base URL to Meta's developer gateway and providing a valid authorization header. Standard chat completion methods accept the full 1 million token context window.

Tool use and autonomous function calling require declaring your schema within the tools array before sending prompts. When a task requires multiple steps, Muse Wonder returns parallel tool calls that your client code must execute and return in sequence. Keep timeouts above 120 seconds for deep reasoning passes, as extensive multi-agent coordination increases latency compared to smaller models.

Active memory operates automatically within supported interactive clients, but custom API implementations require explicit session identifiers. When developing your own wrapper, assign stable thread identifiers to track interactions across days. Always check Meta's developer documentation to verify current rate limits and request limits before scheduling production batch runs.

Performance Benchmarks and Operational Strengths

Meta designed the Muse Wonder series for complex tasks that break standard linear language models. On the vendor-reported SWE-Marathon evaluation, which measures ultra-long-horizon software engineering, Muse Wonder Full scored 57.5 percent at maximum reasoning effort, while Muse Wonder scored 53.1 percent. For domain-specific legal work on Harvey LAB-AA, Muse Wonder Full reached 13.1 percent compared to 10.0 percent for the standard version.

Financial analysis evaluations show competitive results on Vals Finance Agent V2, where Muse Wonder Full achieved 59.1 percent and Muse Wonder scored 56.6 percent. Meta reports that both models retain state across multi-turn queries, scoring above 85 percent on MemoryBench. In end-to-end operational testing on AutomationBench-AA, Muse Wonder Full scored 45.4 percent, demonstrating reliable sub-agent delegation over multi-hour tasks.

Independent verification is still ongoing, and several competing frontier models outperform Muse Wonder on specific evaluations. Claude Opus 5.5 leads on ProgramBench from-scratch coding and HealthBench Pro medical tasks, while GPT-6 Astra leads on the GDP.pdf professional document evaluation. When selecting models for clinical or deep architectural programming, review whether competing lab models better match your error tolerance.

Prompting Techniques and Setup Mistakes to Avoid

Prompting Muse Wonder requires explicit milestone constraints rather than micro-managing step-by-step instructions. Because the model relies on an internal Active Focus mechanism, over-specifying every intermediate action causes the reasoning engine to cycle through redundant self-corrections. Define the final deliverable, specify boundaries for external tool calls, and allow the model to plan its own sub-tasks.

The most common implementation mistake is failing to configure prompt caching when ingesting large repositories or compliance regulations. Sending hundreds of thousands of uncached input tokens repeatedly will exhaust API budgets quickly at $4.00 per million tokens. Storing stable system prompts and document bases at the beginning of the context window reduces input overhead to $0.20 per million tokens.

Another frequent error is assuming that Muse Wonder Full is available through standard API keys. Engineering teams often write backend integration code targeting muse-wonder-full-1.0 only to encounter authorization errors. Build your programmatic production pipelines around muse-wonder-1.0, and reserve Muse Wonder Full for developers working directly inside the Muse Code environment.

Compliance Safeguards and Regulated Workflows

Meta trained the Muse Wonder family under its Superintelligence Scaling Framework to enforce safeguards against chemical, biological, radiological, and nuclear misuse. The architecture also includes guardrails designed to prevent the accidental disclosure of private personal records during multi-step browser navigation. However, Meta notes that a comprehensive formal safety report will only be released in the coming weeks.

This model is not suitable for organizations that require on-premises data isolation or strict zero-data-retention guarantees on day one. Teams subject to Health Insurance Portability and Accountability Act (HIPAA) rules or European Union data residency restrictions must wait for Meta to publish enterprise Business Associate Agreements and sovereign hosting options. If you operate in a strict regulatory environment, use private dedicated clouds instead of public developer APIs.

Our assessment would change if Meta issues enterprise zero-retention commitments and opens the Full model to standalone API access. Until third-party audit reports confirm data boundary enforcement during computer use, regulated firms should restrict live deployments to sandbox environments. In our routine client workflow implementations at Layer3Labs, we advise running parallel validation checks before connecting autonomous agent models to live customer databases.