Meta Muse Wonder: Release Date, Pricing and Specs Explained

Meta released two new AI models on October 12, 2026: Muse Wonder and Muse Wonder Full. They are the first models from Meta Superintelligence Labs that the company describes as "frontier," meaning Meta is positioning them to compete directly with the top-tier models from OpenAI and Anthropic, rather than the lighter, faster Muse Spark line it shipped earlier in 2026. Muse Wonder is available now through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, while the more capable Muse Wonder Full is restricted to paid Muse Code subscribers on the $50-per-month Power Usage plan or higher and is not offered through the API.

What are Muse Wonder and Muse Wonder Full?

Muse Wonder and Muse Wonder Full are Meta's newest large language models, announced on the company's research blog as "a frontier model for personal and professional AI agents." According to Meta's official model card, both models take text, image, video, and audio as input and generate text output, with a shared 1-million-token context window. Meta says the pair represents "the next step on our scaling ladder," built with a larger base model and additional reinforcement learning compared with the earlier Muse Spark family. Muse Wonder Full is the larger, more capable variant of the two; Meta's model card notes it generalizes better than expected to tasks outside its post-training targets, which the company attributes to the strength of its larger base model.

Meta intends the models for commercial and research use, with particular emphasis on agentic work: multi-step planning, tool use, coding, and domain-specific knowledge work in areas like law, finance, and health. The models will also power parts of Meta's consumer Muse app and its artifact-generation features, rolling out first to subscribers.

Release date and availability

Meta published the Muse Wonder announcement on October 12, 2026. Muse Wonder is live today through the Meta Model API (for developers) and through Muse Code, Meta's subscription coding and agent product. Muse Wonder Full is available immediately as well, but only inside Muse Code — Meta has not made it available through its developer API at launch.

Pricing and Muse Code plans

Meta's developer pricing page lists Muse Wonder (model ID muse-wonder-1.0) at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens through the Meta Model API. Muse Wonder Full (muse-wonder-full-1.0) carries no published per-token API rate because it isn't sold through the API; it's bundled instead into two Muse Code subscription tiers.

In practice, that means casual developers can call Muse Wonder through the API at per-token rates comparable to other frontier models, while anyone who wants Muse Wonder Full has to subscribe to Muse Code. Meta frames the Power and Maximum tiers around usage volume, early feature access, and higher file-upload limits rather than separate per-model pricing.

Active Memory and Active Focus: what's actually new

The two headline features Meta highlights for Wonder are Active Memory and Active Focus. Active Memory is designed to retain relevant information about a user across separate conversations and surface it again when it matters for a later task, rather than treating every chat as a blank slate. Active Focus is aimed at agentic persistence: Meta says it helps the model keep working toward a goal by trying alternative approaches, working through blockers, adapting to new information mid-task, asking clarifying questions, and checking in with the user before taking important or irreversible actions.

Meta's model card backs this up with an internal benchmark called MemoryBench, which tests recall accuracy across long multi-turn conversations (64K–100K tokens) that include distractor conversations and conflicting, later-corrected details. A separate internal benchmark, FocusBench, specifically measures how well an agent can persistently work through real-world blockers — things like a failed booking or an unexpected form field — across multiple interactions in domains such as housing, travel, IT support, and personal finance.

Benchmarks and how it stacks up

Meta's model card lists results across three categories: personal and professional agents, software engineering, and general intelligence. For agent-style work, Meta cites its internal Personal Superintelligence Index (a composite of seven evaluations including FocusBench and Agents' Last Exam) alongside third-party results such as Harvey's Legal Agent Benchmark (sourced from Artificial Analysis) and Vals Finance Agent V2 (sourced from Vals AI). For coding, Meta reports results on Terminal-Bench 4.0, DeepSWE v1.1, SWE-Marathon, and ProgramBench, comparing Wonder and Wonder Full against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 at matched "max reasoning effort" settings where possible. Meta is explicit that prompts and tooling for competitor models "may not be optimally tuned," so these are Meta's own comparative claims rather than independently audited results.

Separately, the independent benchmark tracker Artificial Analysis has published its own Intelligence Index scores placing Muse Wonder around 52 and Muse Wonder Full around 56, compared with roughly 58 for Claude Opus 5.5, per reporting from multiple technology outlets. That puts the Wonder family in the same general tier as other current frontier models rather than clearly ahead of them — a different picture from Meta's own benchmark framing, which emphasizes wins on specific agent and memory-focused tasks like MemoryBench and FocusBench rather than a single aggregate score.

Who is Muse Wonder actually for?

Meta's model card frames the intended uses broadly: general-purpose assistance and reasoning, coding and software engineering, personal and professional agents that run multi-step workflows through tool use, domain-specific knowledge work in law, finance, and health, multimodal understanding of documents and video, and research and development support. In practice, that reads as Meta trying to cover the same ground as ChatGPT and Claude for both individual subscribers and developers building agent products on top of the Meta Model API, rather than targeting a narrower niche. Meta also states plainly that the model is not intended for use by people under 18 and that deployers are responsible for assessing and mitigating any risks tied to underage use.

One practical limitation Meta discloses itself: visual and spatial reasoning, especially 3D tasks, "remains a known limitation," and the company recommends developers validate performance themselves for anything depending on precise spatial understanding. Meta also notes that the model hasn't been evaluated across every language present in its pretraining data, and that real-world performance can vary with quantization, serving configuration, and the specific agent harness used to run it — all standard caveats, but ones Meta is upfront about rather than glossing over.

Safety and risk assessments

Meta's model card discloses a formal risk assessment under its own Meta Superintelligence Scaling Framework, covering cybersecurity, chemical and biological (CBRN), and loss-of-control risk domains. Meta states it cannot rule out Muse Wonder Full reaching a "high risk" threshold in the chemical and biological domain without mitigations, and says it has applied safety post-training and system-level classifiers specifically for that domain. With those mitigations in place, Meta reports residual risk at "moderate or lower" across all three domains, which it says supports its decision to deploy.

On adversarial robustness, Meta reports low attack-success rates against jailbreak attempts: single-turn and multi-turn attack success rates of 0.8% on CyberBench, 1.1–2.3% on the cyber subset of PolyGuard, and 8–9% on RedCode, where lower is better. On cyber capability evaluations, Meta reports 88.4% on CyberGym and 58.9% on ExploitBench; for a third evaluation, ExploitGym, Meta notes that while the model captured 92 flags overall, a verification pass found only 21 used the intended vulnerability path, which is the figure comparable to that benchmark's published ranking. On CBRN-specific safety checks, Meta reports refusal rates of 98–100% on BioTIER and an internal chemistry-refusal test, across English, ten language buckets, and adversarially rewritten prompts.

For loss-of-control evaluations, the model card reports an 87.3% score on MASK (which tests whether the model stays honest under pressure to contradict its own knowledge, where higher is better) and a 0% harmful-action rate on an internal Agentic Misalignment evaluation (where lower is better). Meta describes Muse Wonder as its "most aligned model to date" and says a fuller safety report is forthcoming. The company also says it worked with outside groups, including SecureBio, on CBRN-related evaluation benchmarks, and engaged government entities on threat modeling.

How Wonder differs from the original Muse models

Meta's Muse product line started with Muse Spark, a lighter, low-latency model aimed at fast, everyday assistant tasks, alongside separate Muse Image and Muse Video generation models. Wonder and Wonder Full are a step up in scale and ambition: Meta positions them as its first genuine "frontier" release, built on larger base models with more reinforcement learning, and explicitly benchmarked against the leading models from OpenAI and Anthropic rather than against Meta's own earlier releases. The practical difference for users is access and capability tier — Muse Wonder is the new default general-purpose option via the API and Muse Code, while Muse Wonder Full is reserved for paying Muse Code subscribers and is pitched at harder, longer-running agentic work. Readers who want pricing and plan details for Meta's original Muse launch can see Pandromeda's earlier coverage of Meta Muse's plans and features for comparison.

What it means for Meta's AI strategy

This is the first frontier-class model launch since Alexandr Wang took over Meta's AI efforts as part of the restructuring of what became Meta Superintelligence Labs. Shipping a model Meta itself calls frontier — and backing it with a lengthy public risk assessment — signals that Meta wants Wonder judged against the same bar as Claude and GPT-6, not treated as a consumer-feature add-on. The emphasis on Active Memory and Active Focus also tracks with where the rest of the industry has been racing: toward agents that can run longer and remember more, not just answer single prompts well. For comparison, readers can see how Meta's approach to persistent assistants stacks up against how Claude's memory feature works, and how the broader frontier-model race looks against Google's Gemini 4 Argon and OpenAI's GPT-6.

What's next

Meta says Muse Wonder will gradually power more of the consumer Muse app, starting with Muse Code subscribers before a wider rollout, and will also drive the app's artifact-generation features. A full safety report for the Wonder family is still pending, beyond the summary included in the initial model card. Given Meta's stated "scaling ladder" language, further Muse Wonder variants or a broader API rollout for Wonder Full are plausible next steps, though Meta has not announced a timeline for either. For now, developers and consumers evaluating the models should treat Meta's own benchmark claims as self-reported, cross-check them against independent trackers like Artificial Analysis, and watch Meta's model card page for updates as the promised fuller safety report is published.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage