Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, calling them its first frontier-class models. That label matters because Meta has spent more than a year rebuilding its AI effort while OpenAI and Anthropic set the pace. According to TechCrunch and Meta's own announcement, the new models are larger and more capable than the lightweight Muse Spark, with benchmark results in Meta's testing comparable to Claude Fable 5.1 and GPT-6 Astra, and a lead on the MemoryBench test of long-term recall.
The detail builders should notice is not the frontier claim itself, but the split in how you can actually use these models. Muse Wonder is available through the Meta Model API at 4 dollars per million input tokens and 20 dollars per million output tokens, roughly in line with GPT-6 Sol and Claude Opus 5.5. Muse Wonder Full, the stronger version, is not available through the API at all. It is reserved for Muse Code subscribers, including its higher-priced plans, and both models will power Meta's Muse agent, starting with subscribers.
Meta is pitching the pair around two ideas it calls active memory and active focus. In plain terms, active memory is about carrying useful context across conversations and surfacing the right detail for the task at hand, while active focus is about staying on a hard goal, trying another approach when blocked, and checking in before consequential actions. That is a practitioner pitch, not a chatbot pitch. It points at long-running agent work in coding, and at professional tasks in areas like law, finance, and health that Meta says it evaluated.
Both models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year. That matters for a simple reason: this is no longer a research preview assembled from spare capacity. Meta has put dedicated infrastructure, an API price, and a subscription gate behind the same release. Independent testing cited in coverage, including from Artificial Analysis, broadly supports that Wonder is competitive at the top end, even if the exact ranking against rivals such as Gemini 4 Argon remains contested. Treat Meta's own charts as a claim to verify, not a verdict.
For teams building with AI, the practical question is now access, not existence. If you need a programmable frontier model today, Wonder gives you a new API option at a familiar price point, and its memory focus is worth testing on real multi-session workflows rather than single prompts. If you want Meta's best, you have to buy into Muse Code instead of your own stack, which is a deliberate platform choice. Either way, the frontier is now a four-way contest again, and that competition on price, memory, and agent reliability is exactly what builders should want.
