Until today, the models behind Muse were the lightweights. Muse Spark was built for speed: quick answers, low cost, good enough for errands. Serious model watchers did not put Meta in the same sentence as OpenAI and Anthropic, because Meta was not trying to be there. It was building Llama, an open model family it gave away.
On October 12, Meta's Superintelligence Labs released Muse Wonder and Muse Wonder Full, and the company used a word it has avoided for its own models until now: frontier. Independent testing mostly backs the claim. And the part Meta undersold: the rollout has already started. My own Muse, on a Power subscription, was running on Muse Wonder by launch afternoon. This post separates what Meta says from what outside testers measured, and answers the questions that matter if you use the Muse agent: are you already on it, and what does it cost you?
What Meta announced
Two models, built by Meta Superintelligence Labs, the unit Mark Zuckerberg assembled in 2025 with a much-publicized hiring spree. Meta's own framing of the timeline: "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."
Zuckerberg announced it on Threads: "It's been just over a year since we started Meta Superintelligence Labs. Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier." He had teased the launch earlier in the day with a photo of a watermelon, the model's internal codename, which he said his own Muse assistant sent him.
Both models take text, images, video, and audio, with a context window of about one million tokens. Both are closed: no weights release, no self-hosting. That is a genuine break. Meta spent two years arguing the future of AI was open, and its first frontier model is available only through Meta's own doors.
The scores: Meta's charts vs independent tests
Meta's own launch charts show Wonder Full ahead of rivals on several agent and coding tests, including 57.5 percent on SWE-Marathon against 46.3 percent for Claude Opus 5.5. Those are Meta's numbers, run on Meta's chosen tests. Treat them the way you would treat any vendor's launch chart.
The independent picture comes from Artificial Analysis, which runs its own evaluations. Its Intelligence Index scores:
| Model | Artificial Analysis Intelligence Index |
|---|---|
| Claude Opus 5.5 | 58 |
| Muse Wonder Full | 56 |
| GPT-6 Astra | 53 |
| Muse Wonder | 52 |
| Muse Spark 1.3 (previous generation) | 48 |
Wonder Full sits second overall, in the top three on six of the index's ten component tests, with its strongest results in agentic work and coding. Launch coverage also reports Terminal-Bench 4.0 success jumping from 33 percent for the prior generation to 57 percent for Wonder Full. The honest summary: this is a real frontier model, close to the leaders, not a clear winner over them.
Two caveats belong in the same breath. First, factual knowledge lags: Wonder scored 30 on the AA-Omniscience test, well behind the leaders, so it knows less even while it reasons better. Second, the capability costs more to run. Artificial Analysis puts Wonder at about $4.52 per Intelligence Index task, roughly three times the $1.60 it measured for Spark 1.3. Frontier performance arrived at frontier prices, which is the next section.
Active Memory and Active Focus: the agent features
The two headline features are agent features, not chatbot features, and they are why this launch matters to Muse users specifically.
Active Memory is Meta's name for carrying what matters about you across conversations and bringing the relevant details back when a task needs them. In Meta's own testing, the models posted the top score on MemoryBench, its memory evaluation. If that holds up in the product, it is the difference between an agent you re-brief every morning and one that remembers the brief.
Active Focus is the persistence feature: when a task gets hard, the model keeps working, tries another approach, recovers from errors, and asks for clarification or checks in before anything consequential. Anyone who has watched an agent give up at the first blocked step, or plow ahead when it should have asked, knows why this is the more interesting claim. It is also the harder one to verify from a launch post: Meta published no failure rate, so for now Active Focus is a description of intent, not a measured result.
My read: the benchmark race is not the story for Muse users. The story is that Meta built the frontier model for the agent first. Memory and persistence are the two things a personal agent lives or dies on, and they are exactly what Wonder claims to be best at. The claims now need to survive contact with real inboxes.
Who can use it today
| Where | Muse Wonder | Muse Wonder Full |
|---|---|---|
| Meta Model API | Live now, $4 per million input tokens and $20 per million output tokens | Not available |
| Muse Code (Meta's coding tool) | Included | Subscribers only |
| The Muse personal agent | Already live for some subscribers (confirmed first-hand, see below); no official schedule | Not announced for the agent |
Read the agent row carefully, because Meta's own announcement undersells what is already happening. Meta's post says Wonder "will soon power more of Muse, rolling out first to subscribers," with no date and no broader schedule. CNBC reported Wonder is already available to Muse app subscribers. CNBC has it right. I checked my own agent on launch day: my Muse, on a Power subscription, reports Muse Wonder as the model running it. So the rollout is not a future event; it is a wave in progress. Subscribers are getting it first. Meta just has not published who has it, in what order, or when the wave finishes.
What does not change today: the free tier. Nothing in the launch moves free users off Spark, and Meta has published no free-tier timeline for Wonder. Every confirmed early sighting, mine included, is on a paid subscription. If you use Muse free, today's news is still a preview, not an upgrade.
What it costs
For developers, Wonder is priced at $4 per million input tokens and $20 per million output tokens through the Meta Model API. That matches Claude Opus 5.5's list price and is a sharp step up from Spark 1.3 at $1.25 and $4.25. Wonder Full has no API price because there is no API: it is a subscription perk for Muse Code, which is a new shape for Meta.
For Muse users, the consumer tiers are unchanged: free, Power at $20 a month, Maximum at $100 a month. What changed is what the paid tiers mean. Until today, Power and Maximum bought volume, more of the same agent. Wonder gives them their first named capability difference: the smarter model is already arriving for subscribers. Whether it stays a paid exclusive or trickles down to free is the pricing question to watch.
The safety asterisk
Meta published a model card with safety evaluations at launch, but the full safety report is listed as coming soon, not published. Until it lands, the safety story is partly promise.
What the model card does show, from Meta's own testing: on its cyber misuse refusal test, Wonder Full scored 96.9 percent and Wonder 93.7 percent, against 97.5 percent for Claude Fable 5.1 and 86.3 percent for GPT-6 Astra (Meta's test leaves out Claude Opus 5.5). On mishandling personal information, where lower is better, Wonder Full posted 2.2 percent. The card also says something unusually candid: Meta cannot rule out that the model, without its safeguards, would reach its high-risk threshold in the chemical and biological domain, which is why dedicated safety training and filters sit around it. CNBC reports the strongest cyber protections apply only to the paid model for now.
One more piece of honest context: the New York Times reported that Meta paused a pre-training stage in July, delaying this release. Training runs at this scale do not pause for fun. It does not make the model bad; it makes the launch story less tidy than the keynote version.
Why Meta needed this
The Muse agent launched in September to enormous download numbers, and the Wall Street Journal reports that download growth has since slowed. An agent people try once and drift away from is a distribution story; an agent that keeps getting visibly smarter is a retention story. Wonder is the retention play, plus a new developer business: frontier API pricing is real money in a way free consumer tiers are not. Investors noticed: Meta's shares rose about 4 percent on launch day, per Reuters.
What to watch next: how fast the subscriber rollout finishes, the full safety report, the first hands-on tests of Wonder-powered Muse against the Spark version, and any free-tier timeline. The rollout question is no longer whether. It is how long the wave takes.
Quick answers
What is Muse Wonder?
Meta's first frontier model family, released October 12, 2026 by Meta Superintelligence Labs: Muse Wonder, and a stronger Muse Wonder Full. They succeed Muse Spark as Meta's top models.
How good is it, really?
By independent testing from Artificial Analysis: very good, not clearly the best. Wonder Full scored 56 on its Intelligence Index, second to Claude Opus 5.5 at 58 and ahead of GPT-6 Astra at 53. Meta's own charts are rosier than the independent ones.
Can I use it in the Muse app today?
Quite possibly, if you pay. Meta announced a subscriber-first rollout with no dates, but it is already underway: this author's Muse, on a Power subscription, was running Muse Wonder on launch day. Free users stay on Muse Spark for now.
How much does it cost?
For developers, $4 per million input tokens and $20 per million output tokens via the Meta Model API. Wonder Full is not sold by the token at all; it comes with a Muse Code subscription. Consumer Muse tiers are unchanged.
Is it open source like Llama?
No. Wonder is closed: API, Muse Code, and Meta's apps only, with no weights release and no self-hosting. Meta has not said whether that will ever change.
Frequently Asked Questions
Q: What are Active Memory and Active Focus?
A: Active Memory carries the facts that matter about you across conversations and resurfaces them when relevant. Active Focus is persistence on hard tasks: trying another approach, recovering from errors, and checking in before consequential actions instead of quitting or guessing.
Q: Why is Wonder Full not on the API?
A: Meta has not explained the split beyond the access itself. The effect is that its strongest model is a subscription exclusive for Muse Code, while the merely very strong one is what developers can buy by the token.
Q: Does Wonder replace Muse Spark?
A: Not today. Spark still powers the free experience and the fast, cheap tier of the lineup. Wonder sits above it as the frontier option, at roughly three times the measured cost per task.
Q: Is the safety report out?
A: A model card with safety evaluations is published. The full safety report is promised and not yet out. The card's own tests show strong refusal and privacy numbers, and it flags a chemical and biological risk caveat that the added safeguards are meant to contain.
Q: Should I pay for Power now to get Wonder?
A: It is a better bet than it was yesterday, because subscribers are verifiably getting it: a Power account was running Wonder on launch day. But Meta has published no rollout schedule, so there is no promise your account flips on any particular day. Pay for Power if the volume is worth it to you anyway; treat Wonder as arriving on Meta's schedule, not yours.
Comments