Meta introduces Muse Wonder and Muse Wonder Full as its most advanced artificial intelligence models, with a strong focus on coding, autonomous agents, and long-term professional tasks. Initial results show that some of its capabilities are on par with the best competing models, but without establishing overall superiority. The stakes therefore go beyond benchmarks: they also involve persistent memory, controlled access to models, and Meta’s shift in strategy in response to the open ecosystem developed around Llama.
What Just Happened
On October 12, 2026, Meta unveiled Muse Wonder and Muse Wonder Full, two models developed as part of the reorganization of its artificial intelligence operations. The company describes them as its most powerful models to date for personal and professional agent-based tasks.[1]
Their role differs from that of a simple chatbot. Meta cites use cases such as analyzing a lease agreement, understanding an insurance claim, reviewing financial data, or creating a functional application. The models are also designed to work on large software repositories and carry out multi-step operations.
Access, however, remains tiered. Muse Wonder is available in Muse Code and through the Meta Model API. Muse Wonder Full, billed as the most advanced version, is reserved for Muse Code subscribers. Meta also plans to gradually integrate Wonder into its Muse agent, starting with paying users.[1][3]Unlike several generations of Llama, no public download of Muse Wonder’s weights was announced at launch. Meta’s return to the competition for large language models therefore relies, at least initially, on services hosted and controlled by the company.
What’s Really New
The announcement provides neither the number of parameters nor a detailed description of the architecture. It is therefore not possible to determine whether Wonder represents a breakthrough in architecture or is primarily an improvement achieved through pre-training, post-training, increased computational power, and better orchestration of tools.
The most noticeable innovation lies instead in the combination of three capabilities.
The first is the execution of long-running agent-based tasks. Meta claims that Wonder can plan a mission, delegate certain steps to subagents, use multiple tools in parallel, and resume its work after an error.
The second is Active Memory, a memory designed to retain useful information across multiple conversations. This would allow an agent to pick up where they left off on a case without the user having to provide all the background information again.
The third is Active Focus, defined as the ability to stay focused on a goal, try a different approach when facing a roadblock, and seek clarification before taking a major step.[1][3]
These functions should not be attributed solely to the model. An agent’s success also depends on its runtime environment, its tools, its access rights, its memory system, and the rules that determine when it should stop or consult a human.
The documentation lists monthly subscription fees of $50 for the Power Usage plan and $200 for the Maximum Usage plan, both of which include access to Wonder Full. These terms are subject to change and do not yet allow for a precise estimate of the cost of a complete agent-based workflow.[3]
How Does a Model Designed for Agents Work
An agent-based model does not necessarily produce a definitive response after a single instruction. It breaks down the objective into a series of steps: analyzing the request, preparing a plan, searching for information, invoking software, verifying the result, and, if necessary, adjusting its approach.
In a software project, Wonder can, for example, examine multiple files, identify dependencies, modify different parts of the code, and run tests. In a professional setting, it could review documents, extract data, and prepare a deliverable for human validation.
The one-million-token window theoretically facilitates the processing of software deposits or large files. However, it does not guarantee the accuracy of the reasoning or the consistent consideration of each piece of information. A longer context can also increase the cost, processing time, and amount of data exposed to the system.
Persistent memory adds another dimension. It can make the agent more consistent over time, but it requires knowing:
- What information is retained;
- How long it is retained;
- Who can view or edit them;
- How users can correct or delete them;
- If they are reused for other purposes.
Why This Launch Matters for Meta and for Businesses
Muse Wonder marks, first and foremost, a shift in Meta’s strategy. With Llama, the company had heavily emphasized the sharing of weights and the creation of an open ecosystem. Wonder, for now, is focusing on a service-oriented approach, with a standard model accessible via API and a premium version reserved for a proprietary offering.
This development may allow Meta to better control deployment conditions, security safeguards, and monetization. However, it reduces the possibilities for direct auditing, independent hosting, and local adaptation offered by existing weight-based models.
For businesses, the potential value lies not only in the quality of the responses. A model capable of remaining engaged in a long-term task could be used in software development, document analysis, financial research, or the automation of certain processes.
But a professional evaluation must consider the entire system:
- Actual task completion rate;
- Frequency of human validations;
- Average cost per completed assignment;
- Traceability of actions;
- Data protection;
- Reversibility and platform dependency;
- Ability to pause or resume an operation.
A model that performs slightly worse on a benchmark may still be preferable if it offers greater control, more predictable costs, or better integration with the information system.
What Benchmarks Actually Tell Us
The model map published by Meta combines internal evaluations, results provided by developers, and metrics from third-party organizations. The models were generally tested at their maximum reasoning level, which can increase their computational consumption.[2]
These results show that Wonder Full performs competitively on several agent-based and software development tasks. They do not prove that the model is generally superior to GPT-6 Astra, Claude Fable 5.1, or Claude Opus 5.5.
Meta acknowledges that some comparisons use tools, environments, or results from different sources. On SWE-Marathon, for example, competing models are paired with their own coding agents, which limits the strictly controlled nature of the comparison.[2]
The Artificial Analysis Intelligence Index provides a more independent indicator: Wonder Full scores 56, compared to 58 for Claude Opus 5.5. On Terminal-Bench, Wonder Full also trails behind GPT-6 Astra and Claude Opus 5.5. The data therefore supports the idea of Meta’s return among the competitive players, but not that of overall dominance.[5]
Memory, Autonomy, and Data: Issues of Responsibility
An agent with persistent storage and access to applications can process emails, files, calendars, financial information, or confidential business data. An error is no longer limited to an incorrect response. It can trigger an unintended action, a disclosure, or a change that is difficult to undo.
The CNIL and the Council for AI and Digital Technology believe that autonomy, persistent memory, and the ability to interact with multiple services require concrete adjustments to data protection mechanisms.[6] In particular, it is necessary to apply the principle of least privilege, to segregate environments, to limit retention periods, and to require confirmation before irreversible actions are taken.
Meta itself recommends human validation for this type of action. Its model roadmap also acknowledges that Wonder may produce inaccurate or biased responses, make errors in multi-step reasoning, and exhibit variable performance depending on the tools or the agentic environment.[2]
The company also publishes assessments regarding cyber, chemical, and biological capabilities, as well as the risk of loss of control. It estimates that, once its safeguards are in place, the residual risk remains at the “moderate or lower” level in the categories studied. However, this conclusion is based on the company’s own assessment framework. It will need to be supplemented by external testing and by the additional security report announced at launch.[1][2]
In the European Union, the requirements applicable to providers of general-purpose AI models took effect in August 2025. As of August 2, 2026, the European Commission has full authority to enforce compliance with these requirements. Distribution requirements, documentation, systemic risk assessment, and cooperation with downstream actors have therefore become just as important as technical performance.[7]
What to Watch for Now
The scope of Muse Wonder will depend on several factors that are still undecided.
First, we’ll need to verify whether the results can be replicated in independent environments and on real-world professional tasks. Next, we’ll need to measure the total cost of a mission—not just the price of one million tokens—since an agent may make multiple calls to the model and tools.
The development of Wonder Full will also be a key factor. Its limited access currently prevents developers from freely evaluating which version performs best. Its integration into Muse will reveal whether its advertised capabilities can withstand everyday use, imperfect data, and prolonged interactions.
Finally, the decision not to publish the weights raises a strategic question. Does Wonder signal a temporary pause intended to better control a particularly capable model, or a lasting departure from the diffusion model associated with Llama?
Muse Wonder provides reasonable grounds to assert that Meta is rejoining the ranks of labs capable of offering competitive models on certain state-of-the-art evaluations. However, demonstrating a genuine lead will require more than just scores: it will depend on stability in production, cost-effectiveness, transparency, data protection, and users’ ability to maintain control over delegated actions.
To learn more
To understand how Muse Wonder fits into the evolution of boundary models, autonomous agents, and access strategies, check out these analyses from the aivancity blog as well.
Sources
[1] Meta. (October 12, 2026). Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents. View source
[2] Meta. (October 12, 2026). Muse Wonder Model Card. View the model’s profile
[3] Meta for Developers. (2026). Muse Wonder. Documentation accessed October 13, 2026. View documentation
[4] Hart, R. (October 12, 2026). Meta has some Wonder Full new AI models. The Verge. View the article
[5] VentureBeat. (October 13, 2026). Meta’s new Muse Wonder and Muse Wonder Full models achieve state-of-the-art performance, but only the weaker one is available via API. Read the article
[6] CNIL and the Council for AI and Digital Technology. (July 20, 2026). Agent-based AI and Personal Data: Exploratory Note. View the note
[7] European Commission. (2026). General-purpose AI obligations under the AI Act. View source