Meta releases Muse Wonder, its first frontier AI models

Two models, one family. Muse Wonder and Muse Wonder Full are built for what Meta calls personal and professional agentic tasks, meaning jobs where the AI does several steps for you instead of answering one question.

Meta gives everyday examples: navigating an insurance claim, understanding a lease, or making sense of someone's finances. It also points to expert areas like law, finance and health, plus coding for anything from a game to an app for a new business.

TechCrunch describes them as Meta's first attempt at a frontier model. They build on Muse Spark, the lighter model the company launched in April.

• Muse Wonder: available now through the Meta Model API and Muse Code, Meta's tool for writing software with AI.
• Muse Wonder Full: limited to Muse Code subscribers and not served through the API, according to TechCrunch.
• The Muse app: Meta says Wonder will soon power more of Muse, rolling out first to subscribers.

Engadget reports that Wonder Full is open to people on the $50 a month "power usage" and $200 a month "maximum usage" tiers of Muse Code. CNET lists different plan prices, so check Meta's page before you subscribe.

In real life If you pay for Muse today, you will not see Wonder in the app on day one. Developers using the API or Muse Code get it first.

These are the two features Meta wants you to remember, which is fitting.

Active Memory is the model's ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand. Research scientist Kavya Kopparapu said in a launch video that Meta put "a tremendous amount of effort" into remembering what is important to you over days, weeks and months.

Active Focus is the ability to keep working toward your goal when a task gets hard. The model tries other approaches, recovers from errors and follows through. It asks for clarification when needed and checks in before doing anything important.

Why it matters: Long jobs are where agents tend to wander off. Wired notes that these features are meant to keep the model locked onto a stated goal, which matters when an agent researches a topic or builds a large software project.

Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API. TechCrunch says that puts it roughly level with GPT-6 Sol or Claude Opus 5.5.

The Decoder points out that this is a jump from Muse Spark 1.3, which cost $1.25 and $4.25. No token price has been published for Muse Wonder Full.

Artificial Analysis, an independent benchmarking firm, puts the cost of running its Intelligence Index at $4.52 per task for Muse Wonder. Muse Spark 1.3 cost $1.60. Here is how that compares, per The Decoder:

Meta's own tests, shown in its model card, put both models at scores comparable to Claude Fable 5.1 and GPT-6 Astra, TechCrunch reports. Wired adds that Meta says the largest model beats Fable 5.1 and ranks just below Opus 5.5 on the Artificial Analysis Intelligence Index.

The Decoder gives the independent numbers. Muse Wonder Full scores 56 on that index and Muse Wonder 52, up from 48 for Muse Spark 1.3.

The biggest gains are in agent work and coding:

• Terminal-Bench 4.0: the score rose from 33 to 57 percent.
• Coding Agents Index: Wonder Full scores 68, tying Claude Code with Sonnet 5.5, and Wonder scores 63, up from 54.
• ProofBench: Wonder Full solves 94 of 100 math proofs, up from 58 for its predecessor.
• Vals AI Index: Wonder Full debuts in fifth place, and Muse Wonder lands tenth.

Wired also warns that benchmarks are not always a reliable measure of real-world performance.

The catch: Facts. On the AA-Omniscience knowledge test, Wonder Full scores 30 and Wonder 25. Claude Opus 5.5 scores 46, GPT-6 Astra 43 and Gemini 4 Argon 42.

The Decoder reads that gap as a sign the Wonder series sits on a smaller base model than its rivals. Low hallucination rates help: Wonder Full cut its rate from 33 to 23 percent compared with Muse Spark 1.3, second lowest among models with an index score of at least 45. Only Gemini 4 Argon, at 15 percent, is lower.

Finance and legal results stayed at the predecessor's level, dropping up to two points, even though Meta highlights both fields.

Meta was once a trailblazer in open AI. Wired recalls that its Llama models were the first frontier open-weight offering from a big lab, and that the company later fell behind as Anthropic, OpenAI and Google moved faster.

The reboot started 14 months ago, in Meta's words, and CEO Mark Zuckerberg put billions into Meta Superintelligence Labs under Alexandr Wang. Engadget reports that Meta expects to spend as much as $145 billion this year on AI.

Meta says the models were trained at its Prometheus data center in central Ohio. In an interview with the Financial Times, Zuckerberg said that getting here "required a large amount of compute and investment, and a lot of very rigorous scientific work."

An agent that reads your messages, calendar and documents has a lot of access, so Meta spends part of its announcement on guardrails. It says it trained the Wonder family to avoid sharing private information without permission and to turn down requests to help with harmful acts.

Meta also says it kept the models able to help developers and security teams find flaws in their own code. A full report with additional safety evaluations is coming soon, and the release follows Meta's Superintelligence Scaling Framework.

CNET's reporter noted that storing sensitive details raises a privacy question of its own. Zuckerberg told the Financial Times he believes AI can be engineered with normal guardrails and built responsibly. Wired adds that Meta did not say whether agents powered by Wonder misbehaved in testing.

• Developers: you can try Muse Wonder through the Meta API today, at $4 and $20 per million tokens, and compare it with your current model on your own tasks.
• Muse users: expect Wonder to reach the app soon, paying subscribers first, with Active Memory working across your chats.
• Anyone using an AI agent: check what it remembers about you and how to delete it. Memory is the feature, and also the thing to watch.
• Fact-heavy work: keep checking answers against real sources, since knowledge scores trail the top models.