Meta's Superintelligence Lab announced Muse Wonder and Muse Wonder Full on October 12, 2026. TechCrunch describes them as Meta...
Meta's Superintelligence Lab announced Muse Wonder and Muse Wonder Full on October 12, 2026. TechCrunch describes them as Meta's first frontier model, built for personal and professional work that takes planning, reasoning, and several steps to finish.
The pitch is practical. Instead of answering one prompt at a time, these models are meant to handle jobs like an insurance claim, organizing information, or finishing a multi-step task inside your digital tools.
What Makes Muse Wonder Different
Muse Spark, the earlier model in the family, was built around light weight and low latency. Muse Wonder is the bigger one, aimed at demanding tasks, coding, and extended reasoning. Meta calls its approach a scaling ladder toward "personal superintelligence," and the model was reportedly trained at its Prometheus data center in central Ohio.
The focus is on longer work, not short answers: analyzing information, drafting a plan, weighing alternatives, and carrying a task through several stages.
Pricing and Availability
Muse Wonder is available through the Meta Model API and Muse Code, priced at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full goes to Muse Code subscribers rather than through the API. Meta also plans to use the model in its Muse agent.
API access lets developers build the model into their own products, while Muse Code puts it inside coding workflows. Real costs will depend on how much you use it and on Meta's current terms.
Active Memory, Active Focus, and Real Use Cases
Active Memory carries useful context from one conversation to the next, so you repeat yourself less when you pick a task back up. How well that works will depend on the implementation.
Active Focus is meant to keep the model working on hard problems. Meta says it explores alternative approaches, recovers from errors, asks clarifying questions, and checks its own work before important actions. Meta points to FocusBench and SWE-Marathon as evaluations of this behavior.
Meta's demos cover law, finance, health information, and software development, plus a 3D kitchen game, a run plan, merchandise designed from an Instagram profile, and a wedding seating chart. Treat these as examples, not proof of consistency. Legal, financial, and health decisions still need professional judgment and verification.
Context Window and Benchmark Results
The model takes text, image, video, and audio input, returns text, and supports a context window of up to one million tokens. A big window lets it read long material, but that doesn't guarantee it interprets every detail correctly.
Meta's reported evaluations include MemoryBench and comparisons with Fable 5.1 and GPT-6 Astra. Read those carefully. Labs often test with different prompts, tools, and settings, and coding benchmarks in particular rarely compare like for like. If you're evaluating models for a business, weigh benchmarks against accuracy, repeatability, cost, latency, integration effort, and how much human review the output needs.
Safety Evaluations and Known Limitations
Meta reports testing for privacy protection, responses to harmful requests, and cyber capability. The figures include 88.4% on CyberGym, 58.9% on ExploitBench, and 92 flags on ExploitGym (21 from intended vulnerabilities and 71 from unintended ones). Cyber jailbreak attack success ran from 0.8% to 9%. Refusal rates were 98% to 99% for biology-related requests and 99.8% to 100% for chemistry-related ones. Meta also reports 87.3% on MASK and 0% agentic misalignment in the referenced evaluation.
Those numbers describe specific tests. They don't prove the model is safe in general, since results depend on benchmark design, coverage, and threat model, and live systems still need monitoring and safeguards.
Meta is open about the weak spots. The model can be inaccurate or biased, can slip on multi-step tasks, struggles with some 3D spatial reasoning, and performs unevenly across languages. It isn't intended for users under 18, and Meta advises human confirmation before irreversible actions. For organizations, that means clear review steps, careful handling of sensitive data, and authorization for anything consequential.
What This Means for Meta's AI Strategy
The launch ties back to Meta's August 2025 restructuring of its AI work under Alexandr Wang, which the company describes as a ground-up overhaul roughly 14 months before this release. Muse Wonder and Muse Wonder Full are the latest product of that effort.
For developers and businesses, the signal is the shift toward models built for multi-step workflows instead of one-off answers. Whether that pays off depends on reliability, how well the models plug into existing systems, and whether the benefits justify the cost.
If your team is choosing between frontier models for agent-style workflows and wants help testing accuracy, cost, and review requirements before you commit, ATX Soft can help you set up that evaluation.
Frequently Asked Questions
What is Meta Muse Wonder?
Muse Wonder is a frontier AI model from Meta's Superintelligence Lab, built for personal and professional tasks, including complex activities that need planning and multiple steps.
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is described as a more advanced version for Muse Code subscribers and is not offered through the API.
How much does Muse Wonder cost?
Reported pricing is $4 per million input tokens and $20 per million output tokens. Confirm current pricing and terms before deploying.
What are Active Memory and Active Focus?
Active Memory carries context across conversations. Active Focus helps the model work through complex tasks by exploring alternatives, recovering from errors, and asking for clarification when needed.
What tasks can Muse Wonder perform?
Meta's examples include coding, planning, legal and financial work, health-related tasks, game creation, design, and event organizing. Reliability can vary by task.
Does Muse Wonder have safety limitations?
Yes. It can produce inaccurate or biased output, make mistakes on multi-step tasks, and struggle with some spatial reasoning and some languages. Human confirmation is advised before irreversible actions.
Where can users access Muse Wonder?
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.