Muse Wonder: Meta's First Frontier AI Model Is Here

Meta just entered the frontier race for real. On October 12, 2026, Meta's Superintelligence Lab launched Muse Wonder and Muse Wonder Full — the company's first models it calls frontier-class, built on top of Muse Spark. The pitch is agents that remember you: Meta's headline features are Active Memory, which carries what matters across conversations, and Active Focus, which keeps the model working a hard problem instead of giving up. The catch, as always, is the price: Wonder costs $4 per million input tokens and $20 per million output tokens through Meta's API, roughly triple the per-task cost of the Spark model it succeeds. Here is what Meta announced, what the independent benchmarks say, what it will cost you, and the cheapest way to feel the upgrade yourself.

The short answer

Meta launched two models today: Muse Wonder, available now through the Meta Model API and Muse Code, and the stronger Muse Wonder Full, reserved for Muse Code subscribers with no public API. Both are described by Meta as its “most capable models to date,” built for agentic work — software engineering, legal and financial analysis, and long-running tasks that unfold over days, not seconds.

The independent read, from Artificial Analysis (Intelligence Index v4.3.2, max reasoning effort): Wonder Full scores 56 on the Intelligence Index and Wonder scores 52 — a real jump from Muse Spark 1.3's 48 on the same index (v4.3.2), the figure we used in our October 8 ranking. That puts Wonder Full two points behind Claude Opus 5.5 (58) and just ahead of GPT-6 Astra (53). Frontier-adjacent, not the outright leader — which is exactly what “brings us to the frontier,” in Zuckerberg's words, means.

What matters for your wallet: the Muse agent you may already use is about to get this brain. Meta says Wonder “will soon power more of Muse, rolling out first to subscribers.” The price of the upgrade is steep at the API level — but the referral route into the Muse app still carries the 1-billion-token bonus, which is the cheapest seat for the show. Details at the end.

Sources: Meta Newsroom: Introducing Muse Wonder and Muse Wonder Full (official, Oct 12, 2026); TechCrunch: Meta releases its first frontier model, Muse Wonder; Reuters: Meta releases Muse Wonder AI models to challenge Anthropic and OpenAI; VentureBeat: Meta's new Muse Wonder models achieve frontier performance; Artificial Analysis: Muse Wonder and Muse Wonder Full (independent benchmarks); Artificial Analysis: Muse Wonder (max) model page — index 52, $4.52 per task. Primary source: the Meta Newsroom.

What Meta actually announced

From the Meta Newsroom post itself — the primary source, and the only claims you should treat as Meta's own:

• Two models, one family. Muse Wonder and Muse Wonder Full, from Meta Superintelligence Labs (the unit led by former Scale AI CEO Alexandr Wang), “built on the groundwork we laid with Muse Spark.”
• Built for agents, not chat. Meta positions them for “personal and professional agentic tasks” — navigating an insurance claim, making sense of a lease or someone's finances, building an app — plus frontier coding performance.
• Active Memory: “the ability to carry what matters to you across conversations and bring back the details that are relevant to the task at hand.”
• Active Focus: “the ability to keep working toward your goal when a task gets hard, try other approaches, recover from errors, and follow through, asking for clarification when needed and checking in before following through on anything important.”
• Safety framing: trained to avoid sharing private information without permission and to refuse genuinely harmful requests; a full safety evaluation report is “coming soon” — i.e., not published yet.
• Closed weights. Unlike Meta's Llama era, neither Wonder model ships with downloadable weights. This is API-or-nothing, per TechCrunch.

Wonder vs Wonder Full: where you can actually get each

Benchmark scores via Artificial Analysis (Intelligence Index v4.3.2); pricing via TechCrunch/VentureBeat reporting on Meta's materials. The practical upshot: the model posting the headline numbers (Full, 56) is the one you cannot buy by the token. What you can access — Wonder at 52, or the Muse agent as it absorbs the new models — is where the value question lives.

The price of the frontier

This is the part our readers care about, so let's do it in our unit: cost per finished task. Artificial Analysis puts Wonder at $4.52 per benchmark task — nearly triple the $1.60 Artificial Analysis lists for Muse Spark 1.3 on the same index (v4.3.2, the figure in our ranking of October 8). The token math explains it: $4/$20 per million input/output tokens versus Spark's $1.25/$4.25.

That is the eternal frontier trade, stated plainly: Wonder is roughly 8% smarter on the index (52 vs 48) and roughly 180% more expensive per task. Our 7-model ranking keeps showing the same shape — intelligence goes up linearly, cost goes up exponentially. Wonder does not break the curve; it extends it.

Two honest consequences:

• For API builders, Wonder is priced like Opus 5.5 ($4/$20 matches Claude Opus 5.5's list, per TechCrunch) while scoring two index points lower. You are paying frontier prices for near-frontier scores — defensible only if Active Memory wins on your workload. Test on 30–50 of your own tasks before switching anything.
• For everyone else, the Muse app is the arbitrage. Meta monetizes the agent through subscriptions and referral bonuses, not per-token billing. A 1-billion-token sign-up bonus at zero marginal cost beats $4.52/task all day — which is why the CTA at the end of this article exists.

What Active Memory actually means for you

Strip the branding and Active Memory is the feature the whole industry is converging on: the model keeps a working profile of you — your projects, your preferences, the lease you asked about last month — and pulls the relevant pieces back without being re-briefed. Meta's MemoryBench claim (Wonder outscoring Fable, Opus, and Astra on continual-learning recall, per Meta's own testing) is vendor data, but the direction is real: every frontier lab is racing to make memory, not raw reasoning, the differentiator.

Why it matters for money tasks specifically: finance is the canonical long-horizon workload. A tax question in February depends on the freelance income you mentioned in November. An agent that remembers is genuinely more useful than one that is merely smarter per token — and it is also the feature with the sharpest privacy edge. Meta says Wonder is trained to avoid sharing private information without permission. Given that this agent will sit across your messages, calendars, and documents (Meta's own description), read the data terms before you hand it your finances. Memory that works is memory worth protecting.

The honest caveats

• Meta's benchmarks are Meta's. The MemoryBench lead and the “comparable to Fable 5.1 and GPT-6 Astra” framing come from Meta's model card. The independent AA scores (56/52) are the numbers to trust, and they say: excellent, not best.
• The safety report isn't out. Meta promises full evaluations “soon.” A frontier model shipping before its safety report is worth noting, not panicking over.
• Wonder Full is gated. The strongest version has no API and no public price — it's a Muse Code subscriber perk. Most readers will never touch it directly.
• The Llama era is over. Closed weights, API metering, subscriber gates: Meta now plays by OpenAI's rules. The open-source goodwill that built Llama's community does not transfer to Wonder.
• Shares rose 4% (Reuters). The market likes the story. That tells you nothing about the model.

What to do this week

• Don't switch API providers on launch day. Run Wonder against your current model on your own tasks first — 30 to 50 real ones — and compare success rate and cost per finished task, not benchmark scores.
• If you use the Muse agent, watch for the rollout. Wonder powers more of Muse “soon,” subscribers first. The agent getting smarter at no extra per-task cost to you is the actual consumer win here.
• Try it through the bonus, not the API. The Muse referral bonus — 1 billion free tokens for joining with a code — is the cheapest way to feel what the new generation does. The 48-hour redemption clock is real; details below.
• Re-check our ranking next month. When AA publishes Wonder under more evaluations, we'll update the cost-per-task ranking — the $4.52 figure will move as the measurement base widens.

Frequently asked questions

What is Muse Wonder?

Muse Wonder is Meta's first frontier-class AI model family, launched October 12, 2026 by Meta Superintelligence Labs. It comes in two versions: Muse Wonder (API + Muse Code) and the stronger Muse Wonder Full (Muse Code subscribers only). Both target agentic tasks — coding, legal/financial analysis, long-running multi-step work — with features Meta calls Active Memory and Active Focus.

How much does Muse Wonder cost?

Through the Meta Model API: $4 per million input tokens and $20 per million output tokens (cached input $0.20, per VentureBeat). Artificial Analysis estimates $4.52 per benchmark task — nearly triple Muse Spark 1.3's $1.60. Wonder Full has no public API price.

How does Muse Wonder score on benchmarks?

Artificial Analysis scores Wonder Full at 56 and Wonder at 52 on its Intelligence Index (max reasoning effort), versus 48 for Muse Spark 1.3, 58 for Claude Opus 5.5, and 53 for GPT-6 Astra. Meta's own testing claims leads on MemoryBench and parity with Fable 5.1/GPT-6 Astra — vendor data, treat accordingly.

Is Muse Wonder open source?

No. Unlike Meta's Llama models, Wonder ships with closed weights — API and Meta apps only. Meta has not said whether the weights will ever be released.

When will Muse Wonder power the Muse agent?

Meta says Wonder 'will soon power more of Muse, rolling out first to subscribers.' No exact date given. The Muse app remains the cheapest way to access the new generation via the referral token bonus.

Should I switch to Muse Wonder for my API workloads?

Test first. Run 30–50 of your own real tasks on Wonder next to your current model and compare success rate, cost per finished task, and latency. At $4/$20 pricing it matches Opus 5.5's list price while scoring slightly lower — only switch workloads where it clearly wins, and keep a fallback.

Try the Pareto winner free

Muse Spark 1.3 topped our intelligence ranking. Try it with 1 billion free tokens.

Code: S5V9L9

Muse is Meta's AI agent that finishes real errands — email, forms, bookings, research — and keeps working after you close the app. It scored 48 on the Artificial Analysis Intelligence Index, the highest in our 7-model cost-per-task ranking. Join at muse.ai/join, then open Settings > General > Redeem invite code and enter the code above. You get 1 billion free tokens; the person who referred you gets the same.

48-hour deadline: the code must be redeemed within 48 hours of joining, or the tokens are gone for good. Do it the same day.

Claim 1B free tokens with S5V9L9

Full step-by-step: our Muse review.

Apps mentioned in this guide

• Muse review - Delegating real tasks: email, forms, bookings, purcha…

Want the short version for a specific app? Start with our full directory of 13 reviewed tools and apps or see how we rank them.