Meta Muse Wonder: What It Is, Price and How to Try It

Meta's Muse Wonder and Muse Wonder Full went live on October 12, 2026, giving the company its first pair of frontier-class AI models built to go head-to-head with Anthropic's Claude and OpenAI's GPT lineup. Muse Wonder is available now through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is reserved for Muse Code subscribers on the two highest-usage plans. Here's what the models actually do, what they cost, and how they stack up against the rivals Meta is chasing.

What is Meta Muse Wonder?

Muse Wonder is the first release in what Meta calls its "Wonder" series — a pair of models the company describes as its most capable yet for personal and professional agentic work. Meta frames the launch as a step toward "personal superintelligence," and says both models were trained at Prometheus, its gigawatt-scale data center. The series sits under Meta Superintelligence Labs, the research group Meta built out through 2025 to chase frontier-model performance after its earlier Llama-branded releases fell behind Anthropic, OpenAI and Google on independent benchmarks. Meta's own write-up of the launch, published on its AI research blog, lays out the positioning directly: Wonder is meant to be judged against the current frontier, not against Meta's own earlier models.

Unlike the open-weight Llama models Meta has released in the past, Muse Wonder ships as a closed, hosted model family — accessible only through Meta's own API and subscription products, the same distribution model Anthropic and OpenAI use for their top-tier models. That's a meaningful shift in strategy: Llama's whole pitch was that anyone could download the weights and run them anywhere, which is part of why Llama models became a default choice for companies that wanted to self-host rather than depend on a vendor's API. Wonder drops that openness in exchange for keeping Meta's best model under its own control, the same tradeoff Anthropic and OpenAI have made with their own flagship releases. Meta's announcement positions the launch squarely as catching up to rivals rather than open-sourcing a new base model, and doesn't commit to ever releasing Wonder's weights.

Meta's developer documentation lists the full current model lineup, including Muse Wonder, at dev.meta.ai's models reference, which is also where Meta documents each model's context window, available modes and API identifier for developers who want to call it directly rather than through Muse Code or the Muse app.

Muse Wonder vs. Muse Wonder Full: what's the difference

Muse Wonder is the version developers can call directly through the Meta Model API, in Standard mode, with a context window of 1,048,576 tokens. Muse Wonder Full is Meta's more capable sibling, but it isn't available through the pay-as-you-go API at all — Meta's developer documentation confirms it ships exclusively inside Muse Code, the company's subscription coding and agent product, and only on the Power Usage and Maximum Usage tiers.

That split mirrors how Anthropic gates its most capable models behind subscription tiers for some use cases, but it's a departure for Meta specifically, since Wonder Full's unavailability on the API means third-party apps and most developers can't currently build on top of Meta's strongest model — only Muse Code subscribers can use it directly.

How to access Muse Wonder right now

There are three ways to reach the Wonder series today:

• Meta Model API — Muse Wonder (not Wonder Full) is callable by developers as muse-wonder-1.0 in Standard mode.
• Muse Code — Meta's subscription product for coding and agentic tasks offers both models; Wonder Full requires the Power Usage plan or above.
• Muse, the personal agent app — Meta says Wonder will "soon" power experiences inside Muse, rolling out first to existing subscribers, plus artifact generation and other features that need more intelligence. No exact date has been announced for this rollout.

Muse Wonder pricing: API rates and Muse Code plans

Meta's own developer documentation lists per-token API pricing for Muse Wonder in Standard mode:

Meta's Muse Code subscriptions page confirms four named plans — Everyday Usage, High Usage, Power Usage and Maximum Usage — and states that Power Usage and above include "access to the most powerful Muse models, including Muse Wonder Full." That official page does not publish the dollar price of each tier, only which plan unlocks which model. Multiple independent outlets covering the launch, including analyses from AI-pricing trackers citing Meta's own materials, report the tiers at roughly $5, $15, $50 and $200 a month respectively; treat those four figures as reported-but-not-Meta-confirmed until Meta's own pricing page lists them directly, and check Muse Code's own checkout flow for the current price before subscribing. The gap between what Meta documents (which plan gets which model) and what it hasn't yet published (what each plan costs) is worth watching — it's unusual for a company to confirm feature-gating by tier without publishing the tier prices on the same page.

For reference, the full per-token API pricing above comes directly from Meta's own developer pricing and rate-limits page, which is the only place in Meta's documentation where Wonder pricing is stated as a hard number rather than a tier name.

Active Memory and Active Focus, explained

Meta is marketing two features alongside the Wonder models rather than a single flagship capability:

Active Memory is designed to retain context about a user across separate conversations and resurface the details relevant to whatever task is in front of the model next — the kind of persistent memory that previously required a user to re-explain context in every new chat.

Active Focus is aimed at keeping an agent working toward a difficult, multi-step goal: trying alternative approaches when one path stalls, working through blockers, adapting when new information arrives, and — notably — asking the user for clarification or checking in before taking consequential actions, rather than acting autonomously through an entire task chain unsupervised.

Both are squarely aimed at the same agentic-reliability problem Anthropic and OpenAI have been marketing against with their own coding- and task-agents, where models that can technically complete long multi-step jobs still fail by losing the thread partway through, forgetting earlier instructions, or quietly taking an action the user never asked for. Meta's framing of Active Focus specifically calls out "checking in before important actions" as a design goal, which reads as a direct response to the kind of agent incidents that have made headlines this year — cases where an autonomous agent took an unauthorized action, like making a payment or deleting data, because nothing in its design forced it to pause and confirm first.

Meta hasn't published a technical paper detailing exactly how Active Memory stores or retrieves past context, or how long information persists before it's dropped — both are described only at a feature level in the announcement, not at an implementation level. That matters for anyone evaluating the models for real use: "carries context across conversations" can mean anything from a lightweight conversation summary to a persistent profile store, and Meta's current public documentation doesn't say which.

How Muse Wonder's benchmarks compare

Meta's own announcement says the Wonder series achieves state-of-the-art results on benchmarks it names specifically: Harvey's Legal Agent Benchmark (LAB), Meta's own FocusBench, and SWE-Marathon, a long-running software-engineering benchmark. Meta did not publish numeric scores for those benchmarks in the body text of its announcement — only comparison charts — so treat any specific numeric claim you see elsewhere as a third-party reading of Meta's charts rather than a number Meta stated directly in text.

Outside coverage of the launch has cited independent evaluator Artificial Analysis placing Muse Wonder Full close to, but behind, Anthropic's most capable model on its intelligence index — in the same range as the frontier tier that includes Claude Opus 5.5 and Claude Fable 5.1, Anthropic's two current top-end models. On coding specifically, where Meta cites SWE-Marathon, the comparison is to the same class of long-horizon coding agents evaluated in benchmarks like the ones behind GitHub Copilot and Claude Code head-to-head tests. None of these third-party comparisons are Meta's own stated numbers, and Pandromeda has not independently run any of these models — this article reports what Meta and named evaluators have published, not hands-on testing.

Why Meta is pushing into frontier AI now

The Wonder launch comes after a stretch where Meta's publicly released models trailed rivals on independent leaderboards, despite Meta spending heavily to build out Meta Superintelligence Labs and recruit senior AI researchers through 2025. Shifting to a closed, subscription-gated distribution model for its top models — rather than open-weighting them the way Meta did with earlier Llama releases — suggests the company is now prioritizing competing directly on hosted-model revenue and enterprise/consumer subscriptions, the same business model Anthropic and OpenAI rely on, rather than the open-ecosystem strategy that defined Meta's earlier AI positioning.

Meta has also framed the release around safety messaging, calling Muse Wonder its most aligned model to date and pointing to added safeguards against CBRN and cyber-misuse risks, with a fuller safety report promised afterward. That report had not been published as of this writing, which means independent researchers don't yet have Meta's own detailed breakdown of what was tested or how residual risk was measured — only the company's summary claim.

There's also a competitive-pressure angle worth noting: Meta's stock moved on the day of the announcement, which is itself a signal of how closely investors are now tracking frontier-model releases as a proxy for whether a company's broader AI spending is paying off. Meta has spent heavily on data center buildouts like Prometheus and on recruiting senior research talent into Meta Superintelligence Labs since 2025; Wonder is effectively the first public, benchmarked product of that spending that Meta can point to as evidence it's closing the gap with Anthropic, OpenAI and Google rather than falling further behind them.

What to expect next

Three things to watch: whether Meta publishes the Muse Code subscription prices directly on its own site (closing the gap between its documentation and third-party reporting), whether Muse Wonder Full ever reaches the pay-as-you-go API rather than staying subscription-only, and when the promised rollout of Wonder-powered features inside the Muse personal-agent app actually reaches subscribers. Meta's own documentation frames all three as in-progress rather than finished, so expect updates on Meta's developer docs and AI blog rather than a single follow-up announcement.

It's also worth watching whether Meta publishes the full safety report it promised alongside the launch, and whether independent evaluators publish their own benchmark runs against Muse Wonder and Muse Wonder Full rather than relying on Meta's own comparison charts. Until that independent testing lands, the fairest way to read this launch is as Meta's own claim that it has closed the gap with the current frontier — a claim that is plausible given the scale of Meta's investment, but one that hasn't yet been independently verified against the models it's being compared to.

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