Meta's Muse Wonder: An Agent Model That Remembers and Doesn't Quit

The curtain rises: two models, one claim

On October 12, 2026, Meta played for the front row of the AI race for the first time: Muse Wonder and its bigger sibling Muse Wonder Full were introduced. The independent measurement outfit ArtificialAnalysis reports the large model at 56 points in 4th place and the small one at 52. For a company that trailed the pack a year ago, this is a climb from the middle ranks to the top floor: the previous model Spark 1.3 sat at 48. Meta says the new family is built for agent work in daily and professional life: tracking an insurance claim, reading a lease, setting up a working app.

Behind the launch lies a 14-month rebuild. TechCrunch reports the AI team was rebuilt from scratch last year and the models were trained in the Prometheus data center in Ohio at gigawatt scale. Meta, once known for open-weight models, arrived this time with a closed and ambitious family. The company language is grand: one rung up the scaling ladder and a leap toward personal superintelligence. I find the language overblown, but the figures deserve discussion.

Two new skills: remembering and persisting

The first trick of the family is memory. Meta calls it active memory: the model keeps what matters to you for days, weeks and months, and recalls it at the right moment. An analogy: instead of a call center where you retell your story on every call, a corner shopkeeper who knows you. On the company-shared MemoryBench test, Wonder left Fable, Opus and Astra behind. Research scientist Kopparapu strikes the same note: they worked hard on a model that remembers what matters across days, weeks and months.

The second trick is persistence, called active focus: when the work gets hard the model does not quit, it tries another path, asks questions where it gets stuck, and consults you before any important step. Think of the two troubles of current agents: one drops the task halfway, the other goes rogue and does work you never asked for. Meta claims to fix both at once. If true, that is big; but the if matters, and I will come to it.

What it does in daily life

The launch demos flesh out the claim. A traveling user gets an altitude-sickness hint distilled from calendar and mail. In the insurance example, an urgent-care bill, a photo of the insurance card and a dispute text are uploaded; the model drafts the number to call and the words to say, then turns it into a mail draft on request. A wedding seating plan, a T-shirt order from a photo of the cat Coco, a small game called Muse Kitchen: the list goes on. The picture is clear: the model points its eyes at the kitchen and the street more than the office.

The numbers and independent eyes

Now the figures. According to the independent table compiled by The Decoder, Wonder Full ranks in the top three on 6 of 10 tests. Small model at 52, large at 56. Side by side with rivals: Opus 5.5 and Astra ahead, Fable 5.1 left behind, Argon standing out on efficiency. The real jump is against the previous model: Spark 1.3 at 48 points scored 33% on Terminal-Bench 4.0; Wonder Full reaches 57% on the same test while using 30% fewer output tokens. Smarter and thriftier at once.

The software side shines brighter. On the software-agent index the large model ties the rival tool running Sonnet 5.5 at 68 points; the small model at 63 clearly passes the previous version at 54. Mathematical proof test ProofBench climbs from 58 to 94; the leaders sit at 99-100. Wired writes that the Wonder family beats Opus 5.5 and Astra on some software and legal tests. In short, Meta says it counts in software too.

The other face of the coin is knowledge. On the general-knowledge test Omniscience the large model stops at 30 and the small one at 25; Opus 5.5 at 46 and Astra at 43 stand far ahead. The Decoder draws this conclusion: the Wonder family may rest on a smaller base than its rivals. The good news is the hallucination rate: down from 33% to 23%, the second lowest among models above 45 points, with Argon leading at 15%. On safety, Wired notes Meta's refusal training against dangerous outputs and the fact that the company delayed the release for safety.

The legal and finance side is the field Meta truly wants to play on. On the Harvey legal-agent test, Wonder Full at 13.1 points doubles Fable 5.1 at 6.4. The company examples are all knowledge work too: leases, money tables, health information. It makes sense to me: Meta models will enter Facebook, Instagram and glasses, where users ask daily troubles, not software. The model knows less but invents less; a system that stays silent instead of inventing when unsure is worth more in daily work.

What do independent tests say? Engineer Matt Johnston ran the small model through his own tests on a live stream. Mixed results: flashy launch versus real performance differ. Two games scored 88 and 83; in one, enemies take no damage when shot, can be hit with grenades, and the character sometimes clips through walls. A 3D-printer-ready phone case came out clean. Johnston's note is plain: launch day differs from real day.

Developer fal's log is more practical: after watching Wang's promo on X, he set the model up in the terminal, drove it with voice commands, and restyled a site into a dark theme. He also tried native image generation: a portrait blended from two photos came out handsome, but the lettering on the laptop came out broken, small flaws included. The company chart fal flags is bold: on the software-engineering test, Full appears above Opus 5.5 and Astra. Short news channel Curve Wars gives the 74-second summary: 9 of 14 tests sit with Opus, prices tie, these are the company figures at full effort, and no independent judging exists yet.

Price, access and the remaining question

The price side is debated too. Input tokens at $4 per million, output at $20. Roughly three times the previous model: Spark 1.3 stood at $1.25 and $4.25. Cost per task climbed from $1.60 to $4.52; 76% of cost comes from input tokens, and the model uses 28% more input tokens per task. Against rivals: $7.63 on Fable 5.1, $5.98 on Opus 5.5, $3.26 on Astra, $1.99 on Argon, $0.72 on Sol. Costly but not the costliest; The Decoder calls it well placed among similarly priced peers.

Access stays with developers for now. The small model is open through the company model service and Muse Code; the large one for Muse Code subscribers only. It will reach the Muse app soon, paying members first. Plans sit at $50 and $200; Johnston says the $50 plan burned 5 hours of allowance in 15 minutes under parallel tests. The door stands open for the curious, but the tap runs on a meter.

Mashable announced the launch as Meta entering the frontier group for the first time; the outlet notes the family arrived only six weeks after the April Spark. The speed of the trailer shows here: the $14.3B Scale AI move, a new team, a new data center. The company plan is plain too: subscribers first, then specialized smaller models for everyone. The personal-superintelligence talk is grand, but the direction is right: bringing the model down to the assistant in every pocket.

Still, let us keep caution at hand. Every bright figure on the table is the company's own measurement; beyond the index score of the independent outfit ArtificialAnalysis, plenty of items await judging. As Curve Wars puts it, the judges have not scored yet. To the critique that models look pumped on trained tasks, the company answers that it stays strong on unfamiliar work. The knowledge-test lag will be the first place to test that answer.

In my eyes the true exam is not the show but daily life: the person queuing on insurance, parsing a lease clause, seating a wedding. An agent that remembers and refuses to quit makes the difference on those days. Three things to watch in the coming weeks: the arrival of independent scores, the app opening to members, and the price response. Meta said soon for years; this time it says here. Figures will not prove the gap; kitchen results will.