Meta returns to the frontier with Muse Wonder and makes memory the new AI moat
Meta just dropped its first frontier models since last year's restructure. The standout feature is not just raw intelligence, it is memory.
Models · Source: TechCrunch
What happened
Meta just released Muse Wonder and Muse Wonder Full. These are the company's first frontier AI models since Scale AI founder Alexandr Wang restructured Meta's AI division fourteen months ago. They represent a massive leap from the lightweight Muse Spark model. Meta claims these new models rival GPT-6 Astra and Fable 5.1 on benchmarks. The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year.
The real differentiator is a feature Meta calls Active Memory. The models are designed to retain and synthesize user information over days, weeks, and months. Meta research scientist Kavya Kopparapu stated the team put tremendous effort into ensuring the model remembers what is important to the user. Meta says Muse Wonder now scores higher on the MemoryBench benchmark than Fable, Opus, or Astra. The models also feature Active Focus, allowing the AI to recover from errors, try new approaches, and ask for clarification before executing important tasks.
Access is split into two distinct tiers. Muse Wonder is available to developers via the Meta Model API for 4 dollars per million input tokens and 20 dollars per million output tokens. This makes it roughly equivalent in cost to GPT-6 Sol or Opus 5.5. The more powerful Muse Wonder Full is locked behind a Muse Code subscription and will not be served through the API. Both models will soon power Meta's personal AI agent, Muse, rolling out first to subscribers.
Key facts
- $4.00 — Cost per million input tokens for Muse Wonder
- $20.00 — Cost per million output tokens for Muse Wonder
- 14 months — Time since Meta began overhauling its AI efforts
- Prometheus — The Ohio data center where the models were trained
Why it matters
Memory is officially the new battleground for AI agents. Until now, builders focused heavily on raw reasoning and expanding context windows. Meta is shifting the paradigm by prioritizing how well a model remembers user preferences across long timelines. If your AI product treats every session like a blank slate, you are already behind. Active Memory allows agents to handle complex workflows in expert domains like law, finance, and health without needing constant reminders from the user.
The pricing and access strategy reveals Meta's endgame for personal superintelligence. By locking Wonder Full behind a Muse Code subscription and keeping it off the API, Meta is prioritizing its own consumer ecosystem over open developer access. They are building a walled garden for their best technology. Builders relying on Meta for top-tier reasoning via API will have to settle for the standard Wonder model.
Safety and security are also taking center stage as agents get more autonomous. Because personal agents work across messages, calendars, and documents, Meta trained the Muse Wonder family to avoid sharing private information without explicit permission. The models are also trained to improve security and defend against cyber attacks. This shows that frontier models are moving past basic chat and into deep integration with our private digital lives.
For builders
Build long-term agents with Active Memory
Developers can use the Meta Model API to build agents that remember user context over months. This is a massive win for founders building personalized assistants. You can leverage this to create sticky products in healthcare or finance where historical context is everything.
Wonder Full is locked behind a subscription
If you want Meta's absolute best model, you cannot build on it via API. Wonder Full is restricted entirely to Muse Code subscribers. Builders who need maximum frontier performance for their own applications will have to look to competitors like OpenAI or Anthropic.
Use Wonder to find security flaws
Meta specifically trained these models to help developers and security teams find flaws in their own code. Builders can use the standard Muse Wonder model via API to strengthen their systems against cyber attacks. This gives engineering teams a powerful new tool for automated code auditing.
My take
Meta was quiet for over a year, and people thought they were out of the frontier race. They just kicked the door back in. By making Active Memory the core feature, they are telling us that raw intelligence is commoditized and context retention is the real moat. I love that they are focusing on agents that actually remember who you are.
Original reporting: TechCrunch. This is my rewrite and opinion.