Muse Wonder Explained: Architecture, Benchmarks, and Pricing

On October 12, 2026, Meta introduced the Muse Wonder series, releasing Muse Wonder and Muse Wonder Full as frontier artificial intelligence (AI) models built specifically for autonomous personal and professional agent workflows. The release establishes a larger successor family to Muse Spark 1.3, pairing a 1,000,000-token context window with dedicated capabilities for multi-file software engineering, browser navigation, and complex document analysis.

Unlike standard text generation Application Programming Interfaces (APIs) or general chat models such as Claude from Anthropic or ChatGPT from OpenAI, Muse Wonder focuses on persistent execution over extended sessions. Meta paired the model with Active Memory to store user context across separate interactions and Active Focus to enable error recovery and iterative problem solving during multi-hour tasks, rather than stopping when an initial tool call fails.

For technical leads, operations directors, and compliance teams evaluating autonomous agents, Muse Wonder changes the economics and reliability of delegating end-to-end work. The model handles domain-specific evaluation workloads in legal intake, financial analysis, and code refactoring, while providing an OpenAI-compatible API endpoint that integrates directly into existing automation stacks.

What Muse Wonder Is and How the Architecture Operates

Muse Wonder is Meta's frontier model line designed to execute complex, multi-step actions across browsers, desktop software, and code repositories. Meta trained the models at its Prometheus data center following a fourteen-month restructuring of its core AI development programs.

The release splits into two distinct variants with different access paths. Muse Wonder serves as the primary API tier (muse-wonder-1.0), while Muse Wonder Full (muse-wonder-full-1.0) represents the larger variant reserved exclusively for developers on upper-tier coding subscriptions.

Both models feature a 1,000,000-token context window, allowing teams to load entire legal discovery bundles or multi-file repositories into active memory without external retrieval segmentation.

• Active Memory: Stores user profile rules and prior interaction details across separate sessions to reduce repeated prompting.
• Active Focus: Implements autonomous error recovery, pauses to request human clarification before destructive actions, and retries alternate logic paths.
• Advanced Computer Use: Directly controls desktop applications, web browsers, and mobile operating system interfaces to complete administrative forms.
• Parallel Sub-Agents: Plans long-running operations and delegates background tasks to secondary model calls.

Published Benchmark Results across Professional Domains

Meta published benchmark evaluations comparing Muse Wonder and Muse Wonder Full against competing models at maximum reasoning effort. The scores measure performance on specialized legal, financial, and multi-hour software engineering benchmarks.

On the Harvey Legal Agent Benchmark (LAB-AA), Muse Wonder Full scored 13.1 and Muse Wonder scored 10.0, compared to 8.6 for GPT-6 Astra from OpenAI and 4.2 for Claude Opus 5.5 from Anthropic. On Vals Finance Agent V2, Muse Wonder Full scored 59.1, Muse Wonder scored 56.6, and Claude Opus 5.5 scored 58.6.

Software engineering evaluations showed distinct separation on ultra-long-horizon coding tasks. On SWE-Marathon, Muse Wonder Full reached 57.5 under extra-high reasoning, leading Claude Opus 5.5 at 46.3 and GPT-6 Astra at 42.5. However, Claude Opus 5.5 retained leads on ProgramBench from-scratch coding with 70.5 versus 59.0 for Wonder Full, and on Terminal-Bench 4.0-AA with 59.6 versus 57.1.

API Pricing and Commercial Availability Structure

Meta made Muse Wonder available immediately through the Meta Model API and through third-party routing provider OpenRouter. The Meta Model API provides OpenAI Software Development Kit (SDK) compatibility, which allows developers to update their base URL and model identifiers without rewriting request schemas.

Pricing for the Muse Wonder API sits at $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. Prompt caching delivers a 95 percent price reduction on repetitive system prompts and ingested document corpora.

Muse Wonder Full carries no per-token pricing because Meta does not offer it through the standard API. Access requires a Muse Code developer subscription on the Power Usage plan at $50.00 per month or the Maximum Usage plan at $200.00 per month. Consumer access inside Meta's standalone Muse app will roll out to paying subscribers at an unspecified later date.

Target Use Cases and Technical Operational Tradeoffs

Muse Wonder fits workflows requiring sustained context retention and recursive execution over several hours. The model handles complex insurance claim navigation, commercial lease review, and financial reconciliations where an agent must compare disparate records against strict policy rules.

Software development teams can deploy Muse Wonder for multi-file refactoring and migration tickets that cause shorter-horizon models to lose track of dependencies. The Active Focus subsystem allows the model to test its own code, inspect terminal outputs, and adjust logic before submitting pull requests.

The primary operational limitation centers on computer vision and visual user interface control reliability. On the OSWorld V2.1 benchmark, Muse Wonder Full achieved a partial score of 78.4 but a binary completion score of only 39.4, while Claude Fable 5.1 achieved 43.1 binary completion. Teams automating graphical interfaces must maintain human review checkpoints for critical operations.

Safety Frameworks and Regulated Industry Safeguards

Meta trained the Muse Wonder models under the Meta Superintelligence Scaling Framework to prevent unauthorized data exposure and mitigate risks in sensitive operating environments. The system incorporates safeguards against Chemical, Biological, Radiological, and Nuclear (CBRN) information generation while preserving defensive cybersecurity functions.

Meta announced that a comprehensive technical safety report detailing specific red-teaming methodologies will release in the near future. Until third-party compliance audits publish, regulated healthcare and financial institutions must treat raw model outputs as uncertified.

The model actively checks in before completing destructive system actions or external transmissions. Organizations processing sensitive records should combine this internal pause mechanism with client-side data loss prevention gateways to guarantee regulatory compliance.

Implementation Analysis for Regulated Workflow Deployments

At Layer3Labs, we build and run AI systems inside other people's businesses, and we evaluate agent reliability by how models behave when third-party tools return unpredictable errors. In our client intake and onboarding rollouts across legal practices, models without active error recovery pause immediately when a document portal returns an unexpected form field.

Muse Wonder addresses that failure point through its Active Focus loop. In automated document processing, a model that tries alternative DOM selectors before escalating reduces human intervention rates significantly. However, teams that bypass human approval loops on outbound communications face reputational risks if the agent misunderstands nuanced policy boundaries.

Deploying the model effectively requires separating read-only research from external action execution. Review your firm's current API spending on long-context models and establish sandbox testing for Muse Wonder on internal compliance datasets.