A working review of Meta Muse Wonder for creators: what its Active Memory, agentic focus, and 1M context deliver, where its scope stops, and who it fits.
Muse Wonder is Meta's first frontier-class model — a 1M-token context window, an Active Memory system that carries context across conversations, and an Active Focus mode for grinding through hard, multi-step agentic tasks. As a reasoning-and-agent model it is a real leap over Muse Spark; early benchmarks put the larger Wonder Full top-tier on agents, memory, and coding, trailing only Claude Opus 5.5, while the standard Wonder trails both Opus 5.5 and GPT-6 Astra. The catch for creators is unchanged from every other model: it outputs text and plans, not media, and publishes nothing. Score it on agentic work, memory, and coding — not on making or shipping content.
Most Muse Wonder coverage frames it as "Meta finally shipped a frontier model." This review reads it from a creator's seat instead. We build a content engine and read model listings for a living, so the goal is to tell you what Muse Wonder is genuinely good at, where its scope honestly stops, and — because people arrive sideways searching "best AI to make content" — whether Meta's first frontier model belongs in a creator's stack at all.
Short version up top: as a model, this is a real step up. Meta Superintelligence Labs shipped Muse Wonder and the larger Muse Wonder Full on October 12, 2026, calling them its first frontier-class models, and the jump over the Muse Spark line is not incremental. Both carry a 1 million-token context window. The two headline features are built for sustained work: Active Memory retains relevant context across conversations and recalls it when it matters, and Active Focus keeps the model pushing on a hard goal — trying other approaches, recovering from errors, and checking in before important actions. 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, and Meta's own tables show the biggest gains concentrated on agentic benchmarks like SWE-Marathon and Harvey's legal benchmark.
The honest catch is one word, and it did not change from Meta's earlier models: output. Muse Wonder reasons over text and images and drafts well, but it returns words and plans. It generates no video, images, or audio; it renders no branded design; it holds no persistent brand system; and it publishes to nothing. As a reasoning-and-agent layer it is a legitimately strong release. As a content operation it is one step upstream of everything that gets made and shipped.
This review covers what Muse Wonder actually is in 2026, how its memory, agentic focus, multimodal reading, and pricing hold up, where it is the wrong tool, and who should use it versus who should keep looking.
Muse Wonder is Meta Superintelligence Labs' first frontier-class model, positioned for long-running, multi-step agentic work rather than one-shot chat. It ships in two sizes — Muse Wonder and the larger Muse Wonder Full — both carrying a 1 million-token context window, and both trained at Meta's gigawatt-scale Prometheus data center. The two defining capabilities are Active Memory, which retains relevant context across conversations and brings it back when a task needs it, and Active Focus, which keeps the model working toward a hard goal by 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, and describes Muse Wonder as its most aligned model to date. Where this becomes concrete for content is the reading-and-drafting loop: the model is text-first with image inputs, so you can feed it a long source and reference images and ask for an outline, a script, or a repurposing plan, and it drafts one. What it does not do is anything downstream of the words — no media generation, no branded design, no brand governance you configure once, no scheduler, and no publishing. Access, as of October 12, 2026, splits by model: Muse Wonder is on the Meta Model API and in Muse Code (Meta's coding agent) at a reported $4 per million input tokens and $20 per million output; Muse Wonder Full is exclusive to Muse Code subscribers and is not offered via API. Treat the benchmarks and pricing as launch-window figures.
The clearest fit is anyone whose immediate need is agentic reasoning or coding: builders who want a capable model for long-horizon, multi-step tasks that hold context across a thread, and teams that value the Active Memory and Active Focus behaviors for sustained work. It also suits anyone reasoning over long documents or images where the 1M-token window and multimodal reading earn their keep. It is the wrong tool for someone whose actual output is published content — video, images, carousels, scheduled social posts — because producing and distributing that sits entirely outside what a reasoning model does. And it is an awkward fit for a non-technical creator who wants a hosted, make-it-and-post-it product: the standard model is an API and a coding agent, and the strongest variant, Muse Wonder Full, is locked inside a Muse Code subscription with no API at all.
| Dimension | Score | Why |
|---|---|---|
| Agentic tasks & tool use | 4.4 / 5 | The design center — Wonder Full leads rivals on agentic benchmarks like SWE-Marathon and Harvey's legal benchmark, and the model is tuned to sustain multi-step work. |
| Active Memory (cross-session context) | 4.3 / 5 | Retains relevant context across conversations and recalls it when needed — a genuine step up for long, continuous work. |
| Active Focus (persistence on hard goals) | 4.2 / 5 | Tries alternative approaches, recovers from errors, and checks in before important actions instead of giving up or barrelling ahead. |
| Frontier coding | 4.3 / 5 | Strong coding-agent performance; Meta positions it near the top of the field on coding-focused benchmarks. |
| Long context (1M tokens) | 4.2 / 5 | A 1 million-token window on both models, supporting large sources and long agentic sessions. |
| Multimodal reading (image inputs) | 4.0 / 5 | Text-first with image inputs; reasons over photos you feed it, though this is analysis and artifact-building, not media creation. |
| Broad world knowledge | 3.4 / 5 | A relative weak spot for the standard model, which trails Claude Opus 5.5 and GPT-6 Astra on the AA Intelligence Index; Wonder Full trails only Opus 5.5. |
| API pricing / access | 3.3 / 5 | A reported $4/$20 per million tokens is several times the Muse Spark line, and the stronger Full model is Muse Code-only with no API. |
| Media generation (for content) | 1.2 / 5 | Outputs text and plans, not pixels — no image, video, or audio generation, and no branded design. Out of scope. |
| Multi-platform publishing | 1.0 / 5 | No scheduler, no platform integration. It drafts and reasons; it does not post. |
Muse Wonder's pricing tells two stories. The standard model is a reported $4 per million input tokens and $20 per million output on the Meta Model API — a meaningful step up from the Muse Spark line's $1.25/$4.25, which is what you would expect for a bigger, more capable frontier model, but it does raise the stakes of building a high-volume workflow on it. The second story is access: Muse Wonder Full, the stronger variant, is not on the API at all. It is exclusive to Muse Code subscribers, so the best version of this model is reachable only inside Meta's coding agent, not as a building block in a pipeline you own.
The deeper catch is the familiar one for any model: cost per token is not the cost of finished content. The price buys text and plans. Turning that into a published post — the video, the branded card, the scheduling, the multi-platform fan-out — is work and tooling you supply on top, and no per-token price adds media rendering, a brand system, or publishing. For a builder or a heavy drafter, the math is about reasoning quality per dollar; for someone hoping a frontier model is a content shortcut, the token price is the wrong line item entirely.
The honest framing on value: judged as a frontier reasoning-and-agent model, Muse Wonder looks like a strong release worth testing for agentic and coding work, with the caveats that it is pricier than its predecessor and that its best variant is paywalled behind an agent. Judge it against other frontier models, not against a content tool — and confirm current pricing and access on Meta's own pages, since these are launch-window figures.
| Use case | Fit | Why |
|---|---|---|
| Agentic, multi-step tasks that hold context across a thread | Strong | This is what Muse Wonder is built for — Active Memory plus Active Focus are made for sustained, long-horizon work. |
| Frontier coding and tool orchestration | Strong | Strong coding-agent performance, leading rivals on benchmarks like SWE-Marathon and DeepSWE 1.1. |
| Reasoning over long documents or images | Strong | The 1M-token window and image inputs make it capable at analyzing long sources and layouts. |
| Drafting scripts, outlines, and repurposing plans | OK | It drafts well as a raw model, but hands back text you still have to render and publish elsewhere. |
| Keeping copy consistently on-brand at scale | Weak | The raw model has no persistent brand system; Active Memory is thread context, not enforced voice, so drift creeps in across a week. |
| Producing video, images, or carousels for social | Weak | No media generation of any kind. It reasons over images but does not create them. |
| Scheduling and publishing across platforms | Weak | No publishing layer and no scheduler. It drafts and reasons; it does not post. |
| A hosted, make-it-and-publish tool for non-technical creators | Weak | The standard model is an API and a coding agent; the stronger Full variant is subscription-gated inside Muse Code. |
If you arrived at this review wondering whether Muse Wonder can run your content operation, the honest answer is no — and that is a category point, not a knock. The whole thrust of this release is a better brain: more memory, more focus, more reliable multi-step reasoning. That is a real win for anyone building agents. But a content operation is not bottlenecked on reasoning quality — it is bottlenecked on turning an idea into rendered media and getting it published on-brand across platforms. A smarter model does not touch that line item, and the strongest version of this one, Muse Wonder Full, is locked inside Muse Code where you cannot even build on it.
Kompozy sits at that other part of the workflow, and the two are complementary rather than rival. Where Muse Wonder stops at a strong draft or plan, Kompozy turns an idea — or that draft — into 18 content formats: persona and avatar video, carousels, quote cards, infographics, blogs, newsletters, and platform-native posts, held to one brand voice through a Persona Brief and scheduled across eight social platforms plus blog and email. Usefully, Kompozy runs its own generation on managed Claude and OpenAI models — the same frontier class Muse Wonder is competing in — so you get that quality inside the content engine without prompting, copy-pasting, or metering $20-per-million output tokens at all. A practical read: use Muse Wonder when you need to reason, code, or hold context on a hard task, and a content engine when you need to make and publish. Score the model for what it is — Meta's first frontier model — and don't ask a reasoning model to be a content operation.
Muse Wonder is Meta Superintelligence Labs' first frontier-class AI model, launched October 12, 2026 alongside the larger Muse Wonder Full. Both carry a 1 million-token context window and are built for agentic work, with an Active Memory system that carries context across conversations and an Active Focus mode for sticking with hard, multi-step tasks. It is available on the Meta Model API and in Muse Code.
As a frontier reasoning-and-agent model it is a strong release worth testing for agentic and coding work, with a genuine leap over Muse Spark. The caveats: it is pricier than its predecessor, the standard model trails Claude Opus 5.5 and GPT-6 Astra on broad world knowledge, and its best variant is subscription-gated. And it is not worth adopting as a content system, because it generates no media and publishes nothing.
Muse Wonder is the standard model; Muse Wonder Full is the larger, more capable variant that scores higher on early benchmarks. The key practical difference is access: Muse Wonder is on the Meta Model API and in Muse Code, while Muse Wonder Full is exclusive to Muse Code subscribers and is not offered through the API.
No. Muse Wonder is text-first with image inputs — it reasons over images you feed it and builds artifacts from them, but it outputs text and plans, not generated images, video, or audio. To turn what it drafts into finished, published media, pair it with a content engine like Kompozy that renders the media and publishes across platforms.
Reported Meta Model API pricing for Muse Wonder is $4 per million input tokens and $20 per million output — a step up from the Muse Spark line. Muse Wonder Full is not sold by token; it is bundled into a Muse Code subscription. Treat these as launch-window figures and confirm current pricing on Meta's page.
Early benchmarks put Muse Wonder top-tier on agentic tasks, memory, and coding — Artificial Analysis's Intelligence Index scores Wonder Full at 56, tying Claude Sonnet 5.5 and topping GPT-6 Astra (53), with only Claude Opus 5.5 (58) ranking higher. The standard Wonder scores 52, trailing both Opus 5.5 and Astra. The fair read is frontier-class on agents and coding for the larger model, with numbers still early.
Kompozy, if the job is producing and publishing. Muse Wonder reasons, codes, and holds context on hard tasks; Kompozy generates video, images, carousels, blogs, and newsletters and publishes them across platforms. Use Muse Wonder to think and build, and Kompozy to make and ship — and note Kompozy already runs on this class of model under the hood.