Muse Wonder: Meta's First Frontier Model Bets on Memory and Focus

Muse Wonder is Meta's first real swing at the frontier. Meta's Superintelligence Lab released the model alongside a bigger sibling, Wonder Full, on October 12, pitching both as the strongest models it has shipped for personal and professional agents. The base model is live today through the Meta Model API and Muse Code, while the Full version is limited to Muse Code subscribers.

What actually happened

Meta published the launch post on its research blog on October 12. It calls the release "the next step on our scaling ladder" and a "major leap toward personal superintelligence." The post says the team started a ground-up overhaul of Meta's AI stack fourteen months ago and has gone from rebuilding to "pushing the frontier."

TechCrunch frames it as Meta's first attempt at a true frontier model. Muse Spark, the earlier release, was a lightweight low-latency model. The new pair is larger, scores higher on benchmarks and handles heavier tasks like coding.

Availability is split on purpose. The base model is in the API and in Muse Code now, and Meta says it will soon power experiences in the Muse app, rolling out to subscribers first. The bigger Full variant stays inside Muse Code for paying users. VentureBeat's headline sums it up: frontier performance, but only the weaker one is in the API.

Active Memory and Active Focus

Meta is betting the agent war is won on memory. Active Memory, in its words, means carrying important context about you across conversations and pulling up the right details when they matter. TechCrunch says Meta highlighted MemoryBench and claims Muse Wonder now beats Fable, Opus and Astra on it.

"We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months, and beyond," research scientist Kavya Kopparapu said in the launch video, as quoted by TechCrunch.

Active Focus is the persistence half. Meta says the model keeps working toward hard goals, tries alternative approaches, works through blockers and asks for clarification when needed. It checks in before taking important actions. Meta measures this with FocusBench and SWE-Marathon, a test of long-running software projects.

The demos lean consumer. Meta shows the model planning a Saturday run in a browser, ordering custom swag from Instagram photos, building a wedding seating chart and making a playable 3D kitchen game from images.

How it stacks up, per Meta

Every benchmark here is Meta's own. TechCrunch says Meta's model card puts both models in the same range as Fable 5.1 and GPT-6 Astra. Meta also tests domain expertise in law, finance and health, including Harvey's Legal Agent Benchmark.

The more interesting claim is about generalization. Meta plots in-distribution scores against tasks it did not target in post-training, and says the Full model does better on unfamiliar tasks than the Muse Spark 1.3 trend would predict. Its explanation: a bigger, stronger base model carries post-training gains further.

Pricing puts Muse Wonder in premium territory. At $4 in and $20 out per million tokens, TechCrunch calls it roughly equivalent to GPT-6 Sol or Opus 5.5. For context on those rivals, see our GPT-6 Sol and Claude Opus 5.5 write-ups.

Benchmarks from the lab that trained the model are a starting line, not a finish line.

Safety and what is still missing

Meta says this is its most aligned model to date. It claims the strongest safeguards it has shipped against CBRN and cyber misuse, while keeping frontier cyber capabilities for defensive use under its Meta Superintelligence Scaling Framework.

The catch: the full safety report is "available soon," not today. Independent evals have not landed yet either. Until outside testers run it, treat the leaderboard claims as company-reported.

What to do as a reader

If you build agents, the API model is the one you can actually test today. Run Muse Wonder on your own long-memory and long-task workflows before trusting any chart, and compare cost per finished task, not cost per token.

If you are a consumer, watch for the Muse app rollout to subscribers. Memory that follows you across weeks is powerful and also sensitive, so check what you can view, edit and delete.

If you trade AI narratives, Muse Wonder is a big-tech release, not a token. Be skeptical of any coin that borrows the name. Nothing here is investment advice.