Meta has a new frontier model. It’s not open-weight. And it’s not Llama.
Muse Wonder and Muse Wonder Full launched October 12, 2026 — Meta Superintelligence Labs’ first entry in the closed frontier model race. Priced at $4 per million input tokens and $20 per million output, it’s cost-equivalent to OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Opus 5.5. The Full variant is exclusive to Muse Code subscribers and doesn’t have a public API at all.
This is a strategic shift. For years, Meta was the open-source champion — Llama defined the open-weight landscape, and every release was downloadable, fine-tunable, and free. That track continues with Llama (the open-weight series is “paused,” per Meta, not discontinued), but the company is now running two parallel strategies: closed frontier models through Muse, and the open-weight lineage that made Meta the most-used alternative to proprietary models.
The models are built for long-running agentic tasks: coding, legal analysis, finance, computer use, and multi-step workflows. Meta describes the architecture as featuring “Active Memory” and “Active Focus” — proprietary systems that maintain context across extended agent sessions. Technical details remain unpublished, which is standard for closed frontier releases but notable given Meta’s historically transparent approach with Llama.
The timing is significant. Muse Wonder launched six days after Mistral’s Large 4 preview and five days before DeepSeek’s V4.1 Flash update. The October 2026 landscape is fractured: frontier models are either gated (Google’s Gemini 4 Argon is limited to vetted cyber defenders), cancelled (OpenAI’s GPT-6.1 Astra), or positioned as agentic specialists rather than general-purpose intelligence. Meta’s entry adds a fourth category: the closed-but-available frontier competitor.
For the market, the implications are concrete. Teams that wanted Meta’s capabilities in a closed model now have an option — they don’t have to route through Ollama or a third-party inference provider. The pricing puts pressure on OpenAI and Anthropic’s mid-tiers. And the dual-track strategy — closed frontier + open weights — mirrors what Google tried with Gemini and what Anthropic has always done: keep the best models behind the API, let the community build around the rest.
Muse Wonder is available through the Meta Model API and Muse Code. The Full variant remains exclusive. Given Meta’s datacenter expansion completed in Q3 2026, compute availability is no longer the bottleneck it was in 2025. The question now is whether the market wants a third closed frontier option — or whether the open-weight alternative that Llama built will remain the more compelling story.