Meta says Muse Wonder puts it at the frontier, but has published no benchmark scores

Meta has released two new models, Muse Wonder and Muse Wonder Full, and says they put the company at the frontier of AI. The announcement came on October 12, 2026, with Mark Zuckerberg telling the Financial Times that Meta had reached the frontier. For readers searching what Muse Wonder is and whether the claim holds up, here is what is on the record: a product launch, a price list, a safety framework named but not described, and no scores.

That absence is the story. In a race where every lab publishes benchmark tables on launch day, Meta's own announcement leaves them out.

What Meta announced

Meta's post, titled "Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents," calls the pair its "most capable models to date" and says they deliver "state-of-the-art performance on personal and professional agentic tasks." It says they build on Muse Spark and cites strength in expert domains such as law, finance and health, and in frontier coding.

Two features are named. Active Memory is "the ability to carry what matters to you across conversations." Active Focus keeps an agent working toward a goal when a task gets hard, and asks for clarification or checks in before important steps.

Availability is staged. Muse Wonder is available now to developers through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will soon power experiences in its Muse assistant, starting with subscribers.

Price and access

Engadget reports that Muse Wonder costs $4 per million input tokens and $20 per million output tokens. Muse Wonder Full comes with Muse Code's $50 a month and $200 a month tiers. Those figures are Engadget's reading of the launch; Meta's own post does not list prices.

Yahoo Finance reports that Meta says the models can compete with Anthropic's Fable 5.1, OpenAI's GPT-6 Astra and Google's Gemini 4 Argon "in certain benchmarks," without naming the benchmarks or giving scores. We could not find that comparison table in Meta's launch post.

The frontier claim, tested against what is published

Zuckerberg told the Financial Times that the models "allow much more sophisticated behaviors that are at the frontier of what is being done in the AI field," according to Gizmodo's account of the interview. He added: "Getting to this point both required a large amount of compute and investment." Engadget reports that Meta expects to spend as much as $145 billion on AI this year, and that the company describes a rebuild lasting more than a year and tens of billions of dollars.

Money is not capability. A frontier claim needs three things: scores on benchmarks that others can rerun, independent evaluation, and a system card that describes what the model does when pushed. As of October 13, 2026 we could find none of these for Muse Wonder. Neither Engadget nor Gizmodo, both of which we opened, reports any benchmark score or named comparison.

Without those, the only fair comparison is with what Meta has not yet shown. Other labs put benchmark tables in their launch posts so that outside groups such as Artificial Analysis can rerun them within days. If Meta supplies enough for others to test, the claim becomes checkable. Until then it is a statement of intent backed by a spending figure.

The safety side

Meta's post says the models were trained to avoid sharing private information without permission, to refuse requests to perform harmful acts, and to improve security and cyber defenses, while still letting developers find flaws in their own code. It says Meta trains, evaluates and releases models in line with its "Meta Superintelligence Scaling Framework," but gives no details of what that framework contains. A report with additional safety evaluations "will be available soon."

For a lab that argues it is at the frontier, those details are the minimum. The other labs publish system cards on release day. Independent evaluators such as METR and the UK AI Security Institute have been testing models before launch, and the question of which of them has seen Muse Wonder is one Meta has not answered in the materials we opened.

Gizmodo, which takes a skeptical line, also relays reports of serious security flaws found just before launch, including one that could have let users escape Meta's sandbox. We could not verify those reports against a primary source, so we treat them as unconfirmed. Semafor reports that Zuckerberg also rejected calls to slow AI development.

Where it sits in the race

Strategically the launch is a statement of intent. Meta has spent more than a year rebuilding its AI organization under Alexandr Wang, whom it hired through a $14 billion investment in Scale AI, a deal covered in our tracker of Meta Superintelligence Labs. A model that competes in agents is the first public proof that the rebuild is working, which is why the claim is being made loudly.

Our take

Treat Muse Wonder as a serious entry and an unproven claim. The price is competitive on its face, the features are real, and Meta has the money to keep going. But a frontier claim without scores is marketing until someone else supplies them. The sequence we are watching is concrete: the promised safety report, the first independent leaderboard entry, and the first named pre-release evaluator. If all three arrive and the numbers hold, the claim stands. If none does within a few weeks, the launch will have been a pricing announcement.