Meta launches Muse Wonder and Muse Wonder Full with Active Memory, Active Focus and frontier AI performance claims
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, positioning the two models as its most capable AI systems for personal agents, professional work and autonomous software development. Developed by Meta Superintelligence Labs, the new family combines long-context reasoning, persistent memory, computer-use capabilities and mechanisms intended to keep agents working through difficult tasks.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is restricted to eligible Muse Code subscriptions. The standard model costs $4 per million input tokens and $20 per million output tokens, with a 95% cached-input discount. Both models support context windows of up to one million tokens.
The release marks a significant step in Meta's attempt to compete directly with OpenAI, Anthropic and Google at the frontier of AI development. Unlike its earlier Muse Spark models, which emphasized efficient responses, the Wonder family targets sustained execution, including coding projects, browser interactions, document analysis and complex customer-facing workflows.
Meta identifies two capabilities as central to this strategy: Active Memory, which retrieves relevant information from earlier interactions, and Active Focus, which helps agents persist through obstacles, test alternative approaches and request clarification before consequential actions.
The company reports strong benchmark results in coding, professional-domain reasoning and long-running agent tasks. These claims warrant careful interpretation because the evaluation suite includes company-run testing, while independent assessments place Muse Wonder Full among leading models without establishing universal superiority.
··········
MUSE WONDER AND MUSE WONDER FULL TARGET DIFFERENT DEVELOPMENT WORKLOADS.
Meta is introducing two frontier models with different access arrangements rather than offering both through the same public API.
Muse Wonder is the broadly accessible model for developers building applications and autonomous agents. It is available through the Meta Model API and Muse Code, with support for coding, long-running workflows and computer use.
Muse Wonder Full is Meta's higher-capability variant, intended for particularly demanding tasks. At launch, access is limited to Muse Code subscribers on eligible plans, rather than the general Meta Model API.
This distinction matters for organizations evaluating the models. A benchmark advantage demonstrated by Wonder Full cannot automatically be treated as a capability available to developers paying for the standard Wonder API.
Both models support one-million-token context windows, enabling developers to supply extensive documentation, code repositories and task histories. Actual workflow performance still depends on how effectively the agent retrieves relevant information, invokes tools and manages intermediate results.
........
Characteristic | Muse Wonder | Muse Wonder Full |
Positioning | General frontier model | Highest-capability variant |
Meta Model API | Available | Not generally available |
Muse Code | Available | Eligible subscriptions |
Context window | 1 million tokens | 1 million tokens |
API input price | $4 per million tokens | Not published |
API output price | $20 per million tokens | Not published |
Primary workloads | Coding, agents, computer use | Complex reasoning and extended engineering tasks |
........
Meta trained the models using infrastructure at its Prometheus data center, describing the release as the result of improvements to pretraining and reinforcement learning.
The company has not provided enough publicly disclosed architectural information to establish a complete comparison of parameter counts, active parameters, training costs or inference efficiency against competing frontier models.
Consequently, the most useful initial comparisons concern measured task performance, deployment availability and price rather than speculative estimates of model size.
··········
ACTIVE MEMORY RETRIEVES RELEVANT PERSONAL AND PROFESSIONAL CONTEXT ACROSS CONVERSATIONS.
Active Memory is designed to preserve useful information across interactions and retrieve it when a later task requires that information.
Conventional conversational context depends heavily on the material currently supplied to a model. When an interaction becomes long or spans separate sessions, relevant information can be omitted, compressed or lost.
Meta describes Active Memory as a mechanism through which an agent can retain important details and recognize when earlier information becomes useful again.
Consider a personal agent helping someone organize a business trip. The agent may already know the user's preferred departure airport, typical budget and accommodation requirements. When asked to arrange a new trip, it could retrieve those preferences rather than requesting them again.
For professional workflows, an agent might remember the structure of a company's reporting process, previously identified software problems or the conventions used in a particular repository.
The challenge is not simply storing information. Effective memory requires selecting relevant details, retrieving them accurately, identifying when they have become outdated and avoiding inappropriate reuse.
A remembered preference from an earlier conversation should not automatically override a new instruction. Similarly, information associated with one customer or project must not be exposed in another user's session.
Meta highlights performance on memory-focused evaluations, including MemoryBench, as evidence of improvement over competing models.
Those results are relevant to the model's ability to retain and use information under test conditions, but they do not independently establish reliable long-term memory across every application.
Active Memory should also be distinguished from unlimited retention. A one-million-token context window and a system for remembering information across conversations solve related but different problems.
The context window determines how much material a model can process within a supported interaction. Persistent memory concerns what information remains available across separate interactions and how the system decides when to retrieve it.
For enterprise deployments, memory quality must be evaluated alongside data retention policies, access controls, correction mechanisms and deletion requirements.
··········
ACTIVE FOCUS IS DESIGNED TO KEEP AGENTS WORKING THROUGH FAILURES AND MULTI-STEP TASKS.
Active Focus addresses the difference between generating a plausible plan and reliably completing a task that requires sustained execution.
An agent may begin a workflow successfully but encounter unexpected conditions: a webpage changes, an API rejects a request, a test fails or additional customer information becomes necessary.
Meta describes Active Focus as the ability to continue pursuing the original objective, attempt alternative approaches and adapt when circumstances change.
In software development, an agent might generate an initial implementation, execute tests, inspect failures, modify the code and repeat the process until the required behavior is achieved.
For a customer-service task, the agent might need to retrieve an order, determine eligibility for a return, compare replacement options and obtain approval before confirming the transaction.
The central technical challenge is maintaining consistency between the user's objective and the actions taken throughout the workflow.
Long-running execution increases the opportunities for mistakes. An agent can lose track of constraints, repeat unsuccessful actions or interpret information encountered during a task as authorization to change its objective.
Meta says the Wonder family is trained to ask for clarification when necessary and check with users before important actions.
These safeguards are particularly relevant when agents can modify files, submit forms, place orders or interact with financial information.
Meta evaluates sustained task execution through benchmarks including FocusBench and SWE-Marathon, the latter examining performance on extended software-engineering projects.
Strong results on such evaluations indicate progress toward more persistent agents, but production reliability requires additional measures: successful completion rates, intervention frequency, recovery from errors, duplicate actions and adherence to permissions.
Active Focus is therefore best understood as a capability for maintaining progress, not a guarantee that an autonomous agent will finish every task correctly.
··········
META REPORTS FRONTIER PERFORMANCE IN CODING, PROFESSIONAL REASONING AND AGENTIC TASKS.
Meta's benchmark presentation places Muse Wonder and Wonder Full among leading frontier models, particularly in evaluations involving extended execution and professional-domain expertise.
The company emphasizes coding performance across substantial repositories, including tasks requiring planning, parallel modifications, testing and revision.
Wonder Full is positioned as the stronger model for complex engineering work, with Meta highlighting results on DeepSWE and SWE-Marathon.
The company also reports strong performance on specialized evaluations involving law, finance and health.
These categories are commercially important because professional tasks often require interpreting source documents, following detailed instructions and maintaining consistency across multiple stages.
A legal-analysis task may involve examining agreements, identifying relevant clauses and preparing a structured document. Financial analysis may require interpreting statements, reconciling figures and applying specified assumptions.
Such workflows demand more than fluent text generation, but benchmark performance must still be distinguished from professional certification or suitability for unsupervised decisions.
Meta also investigates how improvements on tasks emphasized during training transfer to unfamiliar tasks.
The company reports that Wonder Full generalizes more effectively than a simple extrapolation from its earlier Muse Spark model would predict.
This finding is relevant because AI agents encounter combinations of objectives, tools and constraints that cannot all be represented in training evaluations.
However, generalization remains difficult to quantify through a single benchmark.
Independent comparisons offer a more balanced perspective. In the Artificial Analysis Intelligence Index reported around the launch, Wonder Full achieved approximately 56 points and standard Wonder approximately 52.
........
Model | Reported Intelligence Index |
Claude Opus 5.5 | 58 |
Claude Sonnet 5.5 | 56 |
Muse Wonder Full | 56 |
Gemini 4 Argon | 53 |
GPT-6 Astra | 53 |
Muse Wonder | 52 |
........
These figures represent a particular evaluation snapshot and may change with model updates, reasoning settings and benchmark methodology.
They also show why distinguishing the two Meta variants is necessary. Wonder Full's stronger result cannot be attributed to the standard Wonder model offered through the API.
A model ranking also does not establish the lowest-cost solution for a specific application. Organizations should evaluate representative tasks using the actual model and access tier they intend to deploy.
··········
MUSE WONDER COSTS $4 PER MILLION INPUT TOKENS AND $20 PER MILLION OUTPUT TOKENS.
Meta's published API pricing places standard Muse Wonder in the premium frontier-model category, while Wonder Full currently uses subscription-based access.
Developers using Muse Wonder through the Meta Model API pay $4 per million input tokens and $20 per million output tokens.
Cached input costs $0.20 per million tokens, equivalent to a 95% reduction from the standard input rate.
The supported API model identifier is muse-wonder-1.0.
For developers, the practical cost depends on the proportion of input and output tokens, the effectiveness of caching, the number of tool calls and the frequency of retries.
........
API pricing component | Muse Wonder |
Input, per million tokens | $4.00 |
Cached input, per million tokens | $0.20 |
Output, per million tokens | $20.00 |
Context window | 1 million tokens |
Wonder Full public API price | Not available |
........
For example, an application consuming 100 million input tokens and 20 million output tokens in a month would incur approximately $400 in input charges and $400 in output charges, for a total of $800 in direct model-token costs.
The estimate excludes caching discounts, supporting infrastructure and additional services.
At the same token volume, an application using a model priced at $2 input and $10 output per million tokens would incur $400.
That difference does not automatically establish which model is more economical. A higher-priced model may require fewer attempts, complete tasks more accurately or reduce human supervision.
The relevant metric is the cost of a successfully completed workflow under representative operating conditions.
Wonder Full introduces a separate commercial consideration because access is tied to Muse Code subscriptions.
Meta lists the model under its Power Usage and Maximum Usage plans. The Maximum Usage plan offers substantially higher usage allowances than the Everyday Usage tier, although subscription limits should not be interpreted as unrestricted execution.
Businesses evaluating Wonder Full must therefore consider whether subscription-based access fits their development process, especially when production systems require predictable API throughput and usage-based billing.
··········
COMPUTER USE AND MULTI-AGENT EXECUTION EXTEND THE MODELS BEYOND CONVERSATIONAL ASSISTANCE.
The Wonder family is designed to interact with applications and development environments rather than operate exclusively through text-based exchanges.
Meta describes support for agents that can inspect and operate browsers, desktop software and mobile applications.
These capabilities allow a system to work through interfaces that do not provide dedicated machine-readable APIs.
For example, an agent may need to inspect a dashboard, retrieve information from several applications, update a document and verify that changes were saved correctly.
In coding workflows, Meta supports delegation to sub-agents, parallel tool execution and isolated work environments.
A coordinating agent can divide a large project into smaller tasks, allowing specialized processes to inspect different components or develop changes concurrently.
This can reduce elapsed execution time when tasks are sufficiently independent, but parallel execution introduces coordination problems.
Two agents modifying the same files may produce incompatible changes. An agent working from outdated information may overwrite another agent's results. Tool failures can also leave the overall workflow in an inconsistent state.
Meta's developer materials therefore emphasize isolated worktrees and structured orchestration.
The models are also intended to support artifact generation, including software applications and interactive experiences.
Meta has demonstrated workflows involving the creation of a playable three-dimensional game, personalized merchandise and complex event planning.
These examples illustrate the combination of planning, computer use and content generation, but demonstrations do not establish that every similar task can be completed without intervention.
For production applications, developers need reliable verification mechanisms and clear boundaries around actions that affect external systems.
··········
META PLANS TO INTEGRATE WONDER INTO MUSE, STARTING WITH SUBSCRIBERS.
The launch connects Meta's frontier-model development with its broader strategy of building personal agents capable of carrying out everyday tasks.
Muse is Meta's personal AI agent, designed to work across applications and websites while incorporating relevant user context.
The Wonder models are intended to improve the intelligence available to this system, particularly for complex planning, professional questions and sustained task execution.
However, the October 12 announcement does not mean every Muse user already has access to Wonder.
Meta says Wonder will begin powering additional Muse experiences through a staged rollout, starting with subscribers.
The company also plans to use Wonder to develop specialized models for broader deployment.
This approach separates the most computationally demanding capabilities from potentially more efficient systems intended for high-volume consumer use.
For Meta, the commercial opportunity extends beyond standalone model subscriptions.
An agent that can assist with shopping, travel, communication, professional tasks and everyday administration could become an important interface connecting users with external services.
However, greater autonomy also creates operational and commercial constraints.
Businesses may restrict automated access, require additional authentication or decline to permit particular transactions through third-party agents.
The agent's ability to complete a task therefore depends on more than its reasoning performance. It also requires compatible business systems, authorized access and reliable mechanisms for handling approvals and exceptions.
The recently proposed Personal Agent Protocol, developed by Meta and Sierra with industry partners, addresses related interoperability problems. Nevertheless, its existence does not mean Wonder automatically supports every participating business or that universal agent-to-business integration is already available.
··········
PRIVACY, SECURITY AND USER AUTHORIZATION REMAIN CENTRAL TO PERSONAL AGENT DEPLOYMENT.
A personal agent with access to messages, calendars, documents and applications creates risks that are different from those associated with a conventional chatbot.
Persistent memory can make an assistant more useful, but it also increases the importance of deciding which information should be retained and when it may be reused.
Computer-use capabilities introduce another challenge: information encountered on a webpage or in a document may contain instructions intended to manipulate the agent.
Such prompt-injection attacks can attempt to redirect the system toward unauthorized actions, including disclosing information or modifying files.
Meta says it trained the Wonder family to handle private information carefully, resist adversarial instructions and refuse assistance with harmful activities.
The company also reports additional safeguards for cybersecurity and chemical, biological, radiological and nuclear misuse.
These protections are described within Meta's Superintelligence Scaling Framework.
At launch, Meta indicated that a more comprehensive safety evaluation report would follow.
Until those results are available and independently examined, broad claims of improved safety should be understood as company-reported findings.
For enterprise users, model-level safeguards are only one component of a secure deployment.
Applications should still enforce permissions outside the model, limit tool access, log consequential operations and require appropriate confirmation before financial or irreversible actions.
An agent that remembers user preferences and works persistently toward a goal must also remain capable of recognizing when its authorization is insufficient.
The central requirement is maintaining user control throughout the workflow, including when an agent encounters unexpected circumstances.
··········
WONDER FULL STRENGTHENS META'S FRONTIER POSITION, BUT ITS API RESTRICTIONS LIMIT DIRECT COMMERCIAL COMPARISONS.
The Wonder release represents a substantial improvement in Meta's frontier-model offering, although the distinction between API-accessible Wonder and subscription-only Wonder Full complicates comparisons with competing models.
Wonder Full's reported benchmark performance places it among the strongest systems in the current evaluation landscape.
Standard Wonder offers developers access to advanced coding, memory and agent capabilities through a conventional token-priced API.
These are different products from a deployment perspective.
An enterprise cannot assume that an application using the $4/$20 Wonder API will reproduce Wonder Full's highest benchmark results.
The competitive comparison with Google's Gemini 4 Argon also depends on pricing dates.
Argon's introductory API rates are $2 per million input tokens and $10 per million output tokens through November 30, while its announced December rates rise to $4 and $20.
Wonder's published API pricing therefore matches Argon's announced standard rates rather than its temporary introductory discount.
........
Model | Input / 1M | Output / 1M | Availability distinction |
Muse Wonder | $4 | $20 | Meta Model API |
Muse Wonder Full | — | — | Eligible Muse Code subscriptions |
Gemini 4 Argon, introductory | $2 | $10 | Temporary pricing through November 30 |
Gemini 4 Argon, standard | $4 | $20 | Announced December rates |
Claude Opus 5.5 | $4 | $20 | API access |
GPT-6 Astra | $10 | $50 | API access |
........
The figures compare published or reported token list prices and exclude caching, reasoning-token differences, tool costs and subscription allowances.
The strongest model on a particular benchmark may not provide the best outcome for every application.
An organization building a coding agent should test repository-level tasks, recovery from failures and accepted code changes. A company evaluating professional assistants should examine document accuracy, memory consistency, authorization behavior and review requirements.
For personal agents, the most important measures may be successful completion of real tasks, the number of unnecessary interventions and the frequency of incorrect or unauthorized actions.
Meta's release is especially significant because it attempts to combine these capabilities within a single model family rather than treating memory, reasoning and tool execution as separate features.
Whether that combination creates a durable advantage will depend on how effectively it performs outside controlled evaluations.
Muse Wonder and Muse Wonder Full establish Meta as a stronger competitor in frontier AI, with particular emphasis on persistent memory and sustained agent execution. The next test is whether those capabilities translate into reliable, secure and economically competitive workflows as Wonder reaches more developers and begins powering Meta's Muse assistant.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
