Meta Releases Its First Frontier Model, Muse Wonder, Pitched on "Active Memory" and Priced Like GPT-6 Sol

On October 12, 2026, Meta Superintelligence Labs released two new models: Muse Wonder and Muse Wonder Full. TechCrunch called them Meta's first attempt at a frontier model. The Verge wrote that Meta is making up for lost time in the AI race. Wired argued that the pair puts Meta back at the frontier. The angles differ, but the fact underneath is the same: Meta has shipped something that can sit in one table with the flagship models from OpenAI and Anthropic. Start with the product. Muse Wonder is sold through the Meta Model API at $4 per million input tokens and $20 per million output tokens. TechCrunch notes that this is roughly the price of GPT-6 Sol or Claude Opus 5.5. Muse Wonder Full is the stronger variant. It is limited to Muse Code subscribers and will not be served through the API. Both models will also power Meta's Muse agent. The Verge adds that Meta plans to start with subscribers and later build specialized models "it can bring to everyone using Muse," but it gave no specifics and no timeline.

Now the performance claims, which all come from Meta. According to the model card, the two models reach benchmark scores comparable to Fable 5.1 and GPT-6 Astra. The Verge frames it as competing with GPT-6 Astra, Claude Opus 5.5 and Fable 5.1, and notes that it is unclear how the models compare with Google's Gemini 4 Argon. Wired is more specific. Meta says the largest version beats Anthropic's second-best model, Fable 5.1, on the Artificial Analysis Intelligence Index and ranks just below its best, Claude Opus 5.5. On a few specialized benchmarks, including legal documents and software building, Meta says it beats GPT-6 Astra and both Claude models. Wired is frank that benchmarks are not always a reliable measure of real-world performance. No independent lab has replicated any of this so far.

The most distinctive part of the launch is that Meta sells the model on remembering you. It calls the capability Active Memory, and pairs it with a second feature, Active Focus. Wired describes both as ways to make agents more effective and less prone to misbehave: they keep the model locked on a stated goal so it does not drift during complex research or a large software build, where one slip can compound into expensive mistakes. In the launch video, research scientist Kavya Kopparapu said the team put a tremendous effort into making sure the model can remember what is important to you over days, weeks, months and beyond. On MemoryBench, a benchmark for continual learning and memory construction, Meta says Muse Wonder scores higher than Fable, Opus and Astra.

The background matters too. TechCrunch reports that this is the first frontier release since Scale AI founder Alexandr Wang restructured Meta's AI efforts in August of last year. Since then Meta has shipped the lightweight, low-latency Muse Spark (Wired dates its announcement to April), the Muse agent, the Muse Image model, and a small open-weight model called Muse Glimmer. It says larger open-weight models will follow. Meta, which once led open-weight AI with its Llama series, fell behind after several disappointing versions. Per Meta's post, the new models were trained at its Prometheus data center in central Ohio, which came online earlier this year. Wired notes that Meta did not disclose how the models were built or trained.

Safety is the other thread. Wired reports that Meta says it worked hard to train Muse Wonder to refuse dangerous outputs, such as recipes for bioweapons or help with hacking, and that it scores well on benchmarks of resilience to such prompts. But Wired also points to a run of recent incidents in which agents from Anthropic, OpenAI and others strayed when given open-ended commands. Meta did not say whether agents powered by Muse Wonder misbehaved in testing, or what it did about that specific risk. Zuckerberg wrote on X last month that the company had delayed the Muse release to work on safety, and Meta published a framework for scaling capabilities safely earlier this month. Our reading is that this launch proves comparability, not superiority. Pricing that mirrors GPT-6 Sol and Claude Opus 5.5 says Meta wants to sell a flagship experience. Keeping Wonder Full off the API turns the strongest tier into a subscription moat. Leaning on long-term memory is a deliberate choice of differentiation in the agent race. Watch three things: when a third-party lab publishes independent scores, whether Active Memory holds up in real long-running work, and whether Meta says something more concrete about agent safety. Until those answers arrive, Muse Wonder is a serious claim, not a settled result.