# Meta Releases Muse Wonder, Its First Frontier Model, and Artificial Analysis Puts It Two Points Off the Top

Times of AI Desk · 2026-10-13 · Models

[https://timesof.ai/2026/10/meta-muse-wonder-first-frontier-model](https://timesof.ai/2026/10/meta-muse-wonder-first-frontier-model)

> Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on Oct. 12. Wonder is on the Meta Model API at $4/$20 per million tokens; Wonder Full is for Muse Code subscribers only. Meta says its tests put them alongside Claude Fable 5.1, Opus 5.5 and GPT-6 Astra; Artificial Analysis scores Wonder Full at 56, behind Claude Opus 5.5 at 58. Meta's full safety report is promised but not out.

With **Muse Wonder** and **Muse Wonder Full**, released on Monday, October 12, Meta Superintelligence Labs has for the first time priced and positioned a model as a peer of OpenAI's and Anthropic's top systems, and the first independent score puts it close to them. **TechCrunch** calls it the company's first attempt at a frontier model.

## What shipped

- **Availability:** Muse Wonder is available through the **Meta Model API** and **Muse Code**; Muse Wonder Full is for Muse Code subscribers, Meta says. TechCrunch reports Wonder Full will not be served through the API.
- **Price:** $4 per million input tokens and $20 per million output tokens through the API, which TechCrunch says is roughly equivalent to GPT-6 Sol or Opus 5.5.
- **The agent:** Meta says Wonder "will soon power experiences in Muse," rolling out first to subscribers, and that it will later use Wonder to develop specialized models "that we can bring to everyone using Muse." The Verge notes Meta gave no specifics or timeline for either.
- **Training:** both models were trained at Meta's gigawatt-scale **Prometheus** data center, which TechCrunch places in central Ohio.

## How good, and according to whom

Most of the performance claims are Meta's own. According to Meta's testing, presented in its model card, Wonder and Wonder Full reach benchmark scores comparable to **Claude Fable 5.1** and **GPT-6 Astra**, TechCrunch reports; The Verge says Meta also claims they rival **Claude Opus 5.5**, and that it is unclear how they compare with Google's Gemini 4 Argon. Meta emphasises what it calls **Active Memory**, carrying context about a user across conversations, and **Active Focus**, sustained work on hard goals. On **MemoryBench**, Meta's testing has Wonder ahead of Fable, Opus and Astra, according to TechCrunch.

The independent read came the same day. As of Tuesday morning IST, **Artificial Analysis** lists **Muse Wonder Full (max)** at **56** on its Intelligence Index, tied with Claude Sonnet 5.5 (max) and Claude Opus 5.5 (xhigh) and behind only Claude Opus 5.5 (max) at 58, out of 127 models ranked. That makes it the only non-Anthropic model in the board's top four.

## Context

Meta says it began "a ground-up overhaul" of its AI efforts fourteen months ago; TechCrunch dates Alexandr Wang's restructuring to August of last year. Since then Meta has shipped the lightweight Muse Spark line, most recently [Muse Spark 1.3](https://timesof.ai/2026/09/meta-muse-spark-1-3), and the Muse agent. The agent's rollout has drawn safety reporting, including [404 Media's account of rushed KVM-escape fixes before launch](https://timesof.ai/2026/10/meta-muse-kvm-escape-vulnerabilities-pre-launch) and the New York Times' report that [Zuckerberg called Muse ready to launch despite the risks](https://timesof.ai/2026/10/meta-muse-nyt-zuckerberg-launched-despite-safety-concerns), which Meta disputes. Caixin reported on October 12 that ByteDance and Tencent are fast-tracking personal agents after Muse, as [we reported](https://timesof.ai/2026/10/bytedance-tencent-fast-track-personal-agents-after-muse).

Meta says Muse Wonder has its "strongest safeguards to date" against CBRN and cyber misuse, "while preserving access to frontier cyber capabilities for defensive use," and calls it its "most aligned model to date." The evidence for that is not public yet: "A full report with additional safety evaluations will be available soon," Meta writes.

## Limits

- Meta's benchmark charts did not render in the page text we read, so we quote no per-benchmark scores; comparisons above are as TechCrunch and The Verge characterise Meta's results.
- The Artificial Analysis board shows no date stamp for the Wonder entries; figures are as read on Tuesday morning IST and may change. No human-preference (Arena) result was available.
- The safety report has not been published.

## Sources

- [Meta AI: Muse Wonder: A Frontier Model For Personal & Professional AI Agents (October 12, 2026)](https://research.meta.ai/blog/muse-wonder)
- [TechCrunch: Meta releases its first frontier model, Muse Wonder (October 12, 2026)](https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/)
- [The Verge: Meta has some Wonder Full new AI models (October 12, 2026)](https://www.theverge.com/ai-artificial-intelligence/1009479/meta-muse-wonder-full-ai-model)
- [Caixin Global: ByteDance, Tencent Rush to Launch Personal AI Agents After Meta's Muse Success (October 12, 2026; paywalled, opening only)](https://www.caixinglobal.com/2026-10-12/bytedance-tencent-rush-to-launch-personal-ai-agents-after-metas-muse-success-102492455.html)
- [Times of AI: Caixin: ByteDance and Tencent Fast-Track Personal AI Agents After Meta's Muse](https://timesof.ai/2026/10/bytedance-tencent-fast-track-personal-agents-after-muse)
- [Artificial Analysis: LLM Leaderboard (read October 13, 2026)](https://artificialanalysis.ai/leaderboards/models)
- [Times of AI: Muse Spark 1.3](https://timesof.ai/2026/09/meta-muse-spark-1-3)
- [Times of AI: 404 Media: Meta Rushed Muse KVM-Escape Fixes Before Launch](https://timesof.ai/2026/10/meta-muse-kvm-escape-vulnerabilities-pre-launch)
- [Times of AI: Zuckerberg Called Meta's Muse Agent Ready to Launch Despite the Risks, New York Times Reports](https://timesof.ai/2026/10/meta-muse-nyt-zuckerberg-launched-despite-safety-concerns)