Meta Launches Muse Wonder, Its First True Frontier AI Model
- ai-news
- meta
- muse
- frontier-models
Meta Finally Enters the Frontier Model Race
This week (Oct 12), Meta announced two new models — Muse Wonder and Muse Wonder Full — marking the company's first official entry into the top tier of AI models ("frontier models"), alongside OpenAI, Anthropic, and Google.
The announcement is independently confirmed by three major outlets on the same day — TechCrunch, The Verge, and VentureBeat — underscoring its significance.
What's New?
Until now, Meta focused on its lightweight Muse Spark model and the consumer-facing Muse agent. Now, for the first time, the company is launching larger, more capable models built specifically for complex business tasks: long-running coding work, financial analysis, legal document preparation, and multi-step workflow management.
Independent testing by Artificial Analysis gave Wonder Full a score of 56 on its Intelligence Index — tying Claude Sonnet 5.5, and ahead of GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 (53 each). Only Claude Opus 5.5 (58) scores higher.
Standout Feature: 'Active Memory'
Meta highlighted a new capability called Active Memory — the agent's ability to retain and synthesize personal and professional context across separate conversations over time. According to Meta's internal testing, Muse Wonder leads on the MemoryBench benchmark, ahead of Fable, Opus, and Astra.
Pricing and Access
Wonder is available through the public API at $4/$20 per million input/output tokens — comparable to Opus 5.5 pricing. Wonder Full, the stronger version, is reserved for paying Muse Code subscribers ($50 or $200/month) and is not available through the standard API — a key consideration for businesses needing self-hosted or custom integrations.
What This Means for Businesses
For small and medium businesses considering AI agents for tasks like document analysis, workflow automation, or coding assistance, Meta's entry into the frontier race expands the field of options and will likely push prices and capabilities across the industry. One caveat: Muse Code subscriptions default to a mode that lets Meta use submitted data to improve future models (opt-out available) — worth checking before uploading sensitive information.
Also in AI This Week
Separately, Microsoft announced its third-generation in-house AI chip, Maia 300, promising 40% greater efficiency than leading GPU systems, aimed at reducing Microsoft's reliance on Nvidia hardware. The chip will enter production in 2027, targeting long-running agentic workloads.
Meanwhile, voice AI executives from PolyAI and Otter candidly acknowledged at the HumanX conference that the technology still hasn't reached its 'ChatGPT moment' — a sober reminder that not every corner of AI is advancing at the same pace.