Meta Muse Wonder: Launch, Benchmarks and Pricing (2026)

Last updated: October 13, 2026

Meta is back in the frontier-model race. On Monday 12 October 2026, its Meta Superintelligence Labs (MSL) unit released Muse Wonder and Muse Wonder Full — the company’s first models it is calling frontier-class, and the first serious attempt to catch OpenAI and Anthropic at the top of the charts. Meta calls them its most capable models yet; investors agreed, sending the shares up about 4% on the day (Reuters). But the launch comes with a catch worth understanding before the hype sets in: the strongest model is not the one you can buy.

Quick answer: Meta's Muse Wonder and Wonder Full launched 12 Oct 2026: first frontier models. Wonder Full scores 56 on Artificial Analysis; API costs $4/$20 per 1M tokens.

What Meta actually launched

Two models, built by the lab Meta formed a little over a year ago under former Scale AI chief Alexandr Wang. Where Muse Spark 1.3 — the model behind the Muse personal agent launch in September — was designed as a lightweight, low-latency model, the Wonder pair are larger and aimed at complex work like coding (TechCrunch). Meta says the pair represent “the next step on our scaling ladder and a major leap toward personal superintelligence” (Meta launch post, via TechCrunch). Zuckerberg framed it as the payoff of a roughly 14-month rebuild: “Fourteen months ago, we began an overhaul of our AI efforts, and in that time we’ve gone from rebuilding our foundations to pushing the frontier” (Meta blog post, via Reuters).

The split matters more than the names:

Muse Wonder — the standard model. Available through the Meta Model API (as muse-wonder-1.0) and on OpenRouter (as meta/muse-wonder), priced at $4 per million input tokens and $20 per million output tokens (TechCrunch; VentureBeat).
Muse Wonder Full — the flagship, and the one posting the headline benchmark numbers. It is not served through the API at all. Access requires a Muse Code subscription: the Power tier at $50 a month or the Maximum tier at $200 a month (TechCrunch; subscription tiers reported by Dev and AI Weekly).

Both models carry a 1-million-token context window and accept text, image, video and audio inputs with text output (Dev; Oossa). Meta says Wonder will soon power its Muse personal agent, initially for paying subscribers — and in the Muse app, Wonder rolls out to subscribers first (Reuters; Oossa). If you use Muse through WhatsApp in a supported market, the new model is what your subscription will eventually be paying for; see our guide to using Muse on WhatsApp for where it works today.

Meta Muse Wonder: what the benchmarks actually say

This is where the launch needs reading carefully. Meta’s own charts show Wonder Full beating rivals on several agent and coding tests — but the independent scores are closer. Here is the split, with who measured what:

Benchmark	Wonder / Wonder Full	Best rival	Who measured it
Artificial Analysis Intelligence Index	Wonder Full 56; Wonder 52	Claude Opus 5.5: 58	Artificial Analysis (via VentureBeat) — independent
SWE-Marathon (long coding tasks)	Wonder Full 57.5%	Claude Opus 5.5: 46.3%	Meta’s own charts (via AI Breakfast) — vendor-side
DeepSWE 1.1 (coding)	Wonder Full 79.0	Not stated	Meta model card (via Dev) — vendor-side
Terminal-Bench 4.0	Wonder Full 33% → 57%	—	Reported by Oossa citing Artificial Analysis
MemoryBench (continual learning)	Wonder ahead of Fable, Opus, Astra	—	Meta’s own testing (via TechCrunch); raw numbers not published (AI Weekly)
AA-Omniscience (factual knowledge)	30 points	Rivals score higher	Reported by Oossa citing Artificial Analysis

The honest reading: independently, Wonder Full ties Claude Sonnet 5.5 at 56 and sits ahead of GPT-6 Astra (53), Gemini 4 Argon (53) and Claude Fable 5.1 (53) — but two points behind Opus 5.5 at 58 (VentureBeat, citing Artificial Analysis). The standard Wonder at 52 lands a point behind GPT-6 Astra. Meta’s own charts are stronger — 57.5% on SWE-Marathon against Opus 5.5’s 46.3% — but those are Meta’s runs, and Meta has not published the raw per-benchmark numbers behind its MemoryBench claim (AI Breakfast; AI Weekly). The factual-knowledge score of 30 on AA-Omniscience is the weak spot: the model is built for agents and coding, not for trivia (Oossa).

How much it costs — token price vs job price

The token pricing is simple and exactly Opus-tier: $4 per million input tokens and $20 per million output tokens through the Meta Model API, with cached input reported at $0.20 per million (TechCrunch; VentureBeat). TechCrunch puts that roughly equivalent in cost to GPT-6 Sol or Claude Opus 5.5.

But the price per job tells the sharper story. Artificial Analysis pegs running Muse Wonder at $4.52 per benchmark task — nearly triple the $1.60 per task for the previous Muse Spark 1.3 (Oossa, citing Artificial Analysis). Token prices have jumped too: Spark 1.3 was $1.25 in / $4.25 out, so Wonder’s input price is 3.2× higher and its output price 4.7× higher (Oossa). Frontier capability, frontier bill.

And then there is the Full model. The version with the headline scores cannot be reached with an API key at any price — it is bundled into Muse Code at $50/month (Power) or $200/month (Maximum). Developers comparing Muse Code against the agents in our Muse vs Dots vs Cowork comparison should factor that in: the strongest Meta model is a subscription perk, not a commodity.

Active Memory and Active Focus: what they really are

The two capabilities Meta is marketing hardest both target the same problem: assistants that forget, and agents that drift.

Active Memory carries what the model knows about the user across conversations. “We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” research scientist Kavya Kopparapu said in the launch video (TechCrunch). Meta says Wonder scores higher on MemoryBench than Fable, Opus or Astra — though, again, the launch materials do not publish the raw numbers behind that claim (TechCrunch; AI Weekly).

Active Focus is aimed at multi-step agents: keeping the model working a hard goal through blockers and having it check in before consequential actions, rather than wandering off mid-task (AI Breakfast). It is the direct answer to the agent-reliability problem — and to the safety week the industry just had.

The privacy angle is unavoidable: a model designed to remember you “over days, weeks, months, and beyond” arrives weeks after reporting on the dossiers Meta’s assistant keeps — including on people who never signed up. Our Meta Muse privacy explainer breaks down what is reported, what Meta states, and what remains unverified. Reuters noted the subscription push comes “despite facing concerns over user privacy.”

What Meta is keeping to itself

Three things did not ship with the launch:

Open weights. Neither Wonder model is being released with publicly downloadable weights (VentureBeat). That is a genuine strategic shift: Meta’s earlier Llama models helped popularise downloadable AI that organisations could run on their own infrastructure. The open-weights era at Meta, at least for its flagship, appears over.
The safety report. Meta promised a safety report but had not published it at launch, so the safety claims cannot yet be checked against full test results (Van Data Team).
Broad Muse app access. Wonder reaches the Muse app’s paying subscribers first (Oossa) — free-tier users wait, and availability still varies by market.

Meta says both models were trained at its Prometheus data center in Ohio (AI Breakfast, single source). Its launch demos — a 3D kitchen game, a run planner, custom gifts and a wedding seating chart — are Meta’s own recordings, so treat them as marketing, not benchmarks (Meta demo videos, via the launch coverage).

Who should care about this

Developers choosing an API model: Wonder at $4/$20 is priced exactly like Opus 5.5 but scores two AA index points lower (52 vs 58 for Full-vs-Opus; the API model is the 52). Run your own evals before switching — vendor charts are not your workload.
Teams already paying for Muse Code: Wonder Full at $50–$200/month is now part of the subscription. If you code in that environment, the 57.5% SWE-Marathon number (Meta’s own) is the one to watch, verified against your repos.
Muse app subscribers: Wonder is the upgrade coming to the assistant you already use — with Active Memory, the privacy settings deserve a second look before the rollout reaches you.
Everyone else: this is a signal, not a product yet. The market read is that Meta’s 14-month, Wang-led rebuild has produced a genuine top-five model — and that the company will monetise it through subscriptions and its own tools, not open weights.

Confirmed, reported, unverified

Launch-day coverage moves fast and claims inflate. Here is the status of the key facts:

Claim	Status	Basis
Muse Wonder + Wonder Full launched 12 Oct 2026	Confirmed	Reuters, TechCrunch, VentureBeat same-day reporting
AA Intelligence Index: Full 56, standard 52; Opus 5.5 leads at 58	Confirmed	Artificial Analysis, via VentureBeat
API pricing $4/$20 per 1M tokens; cached input $0.20/M	Confirmed	TechCrunch; VentureBeat
Wonder Full is Muse Code-only, no API access	Confirmed	Meta (via TechCrunch)
Muse Code tiers: Power $50/mo, Maximum $200/mo	Reported	Dev citing Meta’s product page; AI Weekly
1M-token context; text/image/video/audio in, text out	Reported	Dev; Oossa
Closed weights — no public download	Confirmed	VentureBeat
Meta’s SWE-Marathon 57.5% vs Opus 5.5’s 46.3%	Reported (vendor-side)	Meta’s own charts, via AI Breakfast
Trained at Prometheus data center, Ohio	Reported (single source)	AI Breakfast
Safety report promised, not yet published	Reported	Van Data Team
Rolls out to Muse app paying subscribers first	Reported	Oossa

Frequently asked questions

What is Muse Wonder?

Muse Wonder and Muse Wonder Full are Meta’s first frontier-class AI models, launched 12 October 2026 by Meta Superintelligence Labs. They are larger, more capable successors to the lightweight Muse Spark 1.3, aimed at coding and long-running agent tasks, with 1M-token context and features called Active Memory and Active Focus.

How much does Muse Wonder cost?

Through the Meta Model API, the standard Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input reported at $0.20 per million. Wonder Full has no API pricing — it is only available through a Muse Code subscription (Power $50/month or Maximum $200/month).

Can I use Muse Wonder Full via API?

No. Meta has confirmed Wonder Full is not served through the API; it is reserved for Muse Code subscribers. Only the standard Muse Wonder is available via the Meta Model API and OpenRouter.

Is Muse Wonder open source?

No. Neither Wonder model ships with publicly downloadable weights — a marked shift from Meta’s Llama era, when downloadable models were its signature.

When will Muse Wonder come to the Muse app?

Meta says Wonder will soon power its Muse personal AI agent, initially for subscribers, and reporting indicates paying subscribers get it first. Availability varies by market — Muse itself is still limited to the US, Canada and select regions.

Sources and methodology

This article was desk-researched on 13 October 2026 from same-day launch coverage, separating vendor claims from independent measurement. Claims rest on: Reuters’ launch report (Meta releases Muse Wonder AI models to challenge Anthropic and OpenAI, 12 Oct 2026); TechCrunch (Meta releases its first frontier model, Muse Wonder, 12 Oct 2026); VentureBeat (Meta’s new Muse Wonder and Muse Wonder Full models achieve frontier performance, 12 Oct 2026); and AI Weekly’s launch alert (Meta releases Muse Wonder, its first frontier-class AI model). Benchmark figures from Meta’s own materials are labelled vendor-side; independent figures come via Artificial Analysis as reported by VentureBeat, Oossa and AI Breakfast. Nothing here is hands-on tested — API access was not exercised for this article. Correlations and vendor claims are labelled as such throughout; no causal claims are made beyond what the cited sources state.

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!