Meta Superintelligence Labs released a pair of models it calls frontier-class — a 1M-token context window, an Active Memory system that carries context across conversations, and an Active Focus mode for sticking with hard, multi-step tasks — with the larger Full model locked behind a Muse Code subscription.
2026-10-12 · by Moe Ameen
Meta released Muse Wonder and Muse Wonder Full on October 12, 2026 — the first models it is willing to call frontier-class. They come from Meta Superintelligence Labs and build on the Muse Spark family that preceded them. Meta says both were trained at its gigawatt-scale Prometheus data center, with work spanning pretraining and reinforcement learning, and describes Muse Wonder as its most aligned model to date with its strongest safeguards yet against CBRN and cyber misuse.
The design center is agentic work, not one-shot chat. Both models carry a 1 million-token context window, and the two headline capabilities are built for long-horizon tasks: Active Memory retains relevant context across conversations and brings back details when they matter, and Active Focus keeps the model working toward a hard goal — trying alternative approaches, recovering from errors, and checking in before important or irreversible actions. Meta cites domain expertise in law, finance, and health, frontier coding, and the ability to work in a browser and across apps. On modality, the models are text-first with image inputs; Meta's demos build artifacts from photos — designing custom merchandise from images in an Instagram profile, and generating a 3D game from image inputs — and it says Muse Wonder will soon power experiences in Muse, rolling out to subscribers first.
Access splits the two models. Muse Wonder is available through the Meta Model API and through Muse Code (Meta's coding agent), with reported API pricing of $4 per million input tokens and $20 per million output tokens — a step up from the Muse Spark line. Muse Wonder Full, the larger model, is exclusive to Muse Code subscribers and is not offered through the API, which makes this Meta's first frontier model to sit behind a paywall rather than ship openly.
The benchmark picture is early and leans on Meta's own testing plus third-party evaluators. On Artificial Analysis's Intelligence Index, Muse Wonder Full scores 56 — tying Claude Sonnet 5.5 and topping GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 (53 each) — while the standard Muse Wonder scores 52, trailing GPT-6 Astra on that measure; only Claude Opus 5.5 (58) ranks above Wonder Full. Meta's own benchmark tables show Wonder Full leading rivals on some agentic evals, like SWE-Marathon and Harvey's legal benchmark, while trailing on others, like Terminal-Bench — so the honest read is strong but mixed on agents and coding, not a clean sweep, and still behind on broad world knowledge for the standard model. Treat the specific scores and pricing as launch-window figures and confirm current numbers on Meta's own pages.
The most creator-relevant line in this launch is the one about Active Memory and Active Focus — a model that holds context across a long session and grinds through a hard, multi-step goal. That reads like a description of a publishing pipeline, and it is the tell: the frontier is getting good at sustained, reliable, multi-step work, which is exactly the discipline that separates a demo from a content operation. But a model that remembers your thread and clicks through a browser is still not a content engine. It generates no video, renders no branded design, holds no brand system you configure once, and publishes to nothing — and the strongest version of it is locked inside Muse Code, not even reachable by API. The practical move is not to subscribe to a coding agent and hand-build a content stack on top of a pricier model. It is to use an engine already built for the job.
Kompozy is that engine, and it is model-agnostic by design — its own generation runs on managed Claude and OpenAI models, the same frontier class Muse Wonder is competing in, so the quality is already inside the product with no API to wire and no $4/$20-per-million meter to watch. Drop an idea or a Muse Wonder draft into Kompozy and it fans into a Persona Short or a HeyGen avatar video with a face-locked identity, Clipped Shorts from long-form, brand-exact Carousels and Quote Graphics, a Blog Article, an Email Newsletter, and native Text Posts — all held to one voice by your Persona Brief, then scheduled across eight social platforms plus blog and email on Autopilot, with a per-post review pipeline gating every output before it ships. Muse Wonder is a stronger brain; Kompozy is the steady hand for the one job the brain doesn't do — making the content and getting it live.
They are Meta Superintelligence Labs' first frontier-class AI models, launched October 12, 2026. Both carry a 1 million-token context window and are built for agentic work, with an Active Memory system that retains context across conversations and an Active Focus mode for sticking with hard, multi-step tasks. Muse Wonder is the standard model; Muse Wonder Full is the larger, more capable variant.
Muse Wonder is available through the Meta Model API and through Muse Code, with reported pricing of $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is exclusive to Muse Code subscribers and is not offered through the API. Treat the pricing as a launch-window figure and confirm current numbers on Meta's own pages.
No. Muse Wonder drafts text and reasons over images you feed it, but it returns words and plans in an agent session or an API response. It does not render video or carousels, hold a brand voice across a week, size posts per platform, or schedule and publish anywhere. A content engine like Kompozy turns its drafts into finished, on-brand formats and publishes them across platforms.
Early benchmarks put Muse Wonder top-tier on agentic tasks, memory, and coding — Artificial Analysis's Intelligence Index scores it 56 (Full), which ties Claude Sonnet 5.5 and tops GPT-6 Astra, and 52 (Wonder), which trails Astra. Only Claude Opus 5.5 (58) ranks above Wonder Full on that index. The fair read is frontier-class on agents and coding, with the numbers still early.