Meta-chan heads for the front row. The new student gains the power of "Wonder".
Meet Meta-chan (Meta AI), who only enrolled on October 4, 2026. On October 12 (US time), her family Meta announced "Muse Wonder" โ the first model it describes as having reached the frontier (the cutting edge) โ along with her big sister, "Muse Wonder Full". What can they do, where can you use them, and how do they differ from Kuroko and friends? Here's a gentle guide.
So what is Muse Wonder, anyway?
The biggest change: from "light and quick" to "big and persistent." Muse Spark, the model so far, is light and fast to respond. Wonder is larger, built for hard work like coding and long-running agent tasks (TechCrunch). Mark Zuckerberg described it as a much larger, more advanced kind of model โ one that takes Meta to the frontier.
โก Light and quick
- Debuted on April 8, 2026 as MSL's first model
- Moved up a grade three times from summer into fall: 1.1, 1.2, 1.3
- Sparkย 1.3 API: input $1.25 / output $4.25 (per 1M tokens)
๐ Two big sisters
- Younger sister Muse Wonder (API ID muse-wonder-1.0) and her even bigger older sister Muse Wonder Full
- Reads text, images, video and audio; answers in text. Context window: 1M tokens
- Closed: the weights are not handed out. Meta moved away from the Llama-style "give it away" approach with Muse Spark in April, and has shown no plans to release Wonder's weights either
Active Memory
A new selling point: it remembers what matters across conversations, carrying important details from one chat into the next.
Active Focus
When stuck, it tries another approach, and it checks with you before important actions โ so long errands don't get abandoned halfway.
Codename: "Watermelon"
Its internal codename during development was Watermelon, as revealed by MSL chief Alexandr Wang, according to CNBC.
Trained at "Prometheus"
Trained at a gigawatt-scale data center in Ohio. Note that it is not intended for users under 18.
Where can you use them?
Here's the map as of October 12, 2026. The key point: the two sisters come in through different doors.
Paid API (Meta Model API)
Only the younger sister, Wonder. Per 1M tokens: input $4 / output $20 (cached input $0.20). Wonder is not offered in the cheaper Contributor mode, where you pay with your data.
Muse Code
In the coding tool Muse Code, Wonder is on every plan. Big sister Wonder Full is only on the top Power and Maximum plans, and is not on the API. Monthly prices are not listed on the official page; CNBC reports Wonder Full comes with the $50 or $200 per month plans.
The Muse app
For the personal agent Muse, the official line is that Wonder will soon power the Muse experience, starting with paid subscribers. At announcement, it did not say the app already uses it.
What about Japan?
The Muse app launched in the US on September 8 and in Canada on September 18. Nothing has been announced about availability in Japan, Wonder included.
In Muse Code, training is on by default
When you subscribe to Muse Code, Contributor mode is switched on from the start. Your code and prompts are used to train Meta's models, and in return you get more usage. The data-science site letsdatascience (October 13) criticized this default. If you work with code from your job, it's wise to check the setting right after you sign up.
The Wonder sisters' strong subjects
All numbers are self-reported, according to Meta's announcement (model card). The comparison models are Claude Fableย 5.1 and Claude Opusย 5.5 from Kuroko's family and GPT-6ย Astra from Chappy's family, all at maximum settings.
Remembering (memory)
MemoryBench (Meta's internal test): Wonderย 85.6, Fullย 85.5. Opusย 5.5ย 81.5, Astraย 81.4, Fableย 5.1ย 79.0.
Long coding tasks
DeepSWE 1.1: Fullย 79.0, Wonderย 78.8. Opusย 5.5ย 74.2, Astraย 74.1, Fableย 5.1ย 67.4.
Ultra-long projects
SWE-Marathon: Fullย 57.5 (at xhigh), Wonderย 53.1. Opusย 5.5ย 46.3, Fableย 5.1ย 45.6, Astraย 42.5.
Business automation
AutomationBench-AA: Fullย 45.4. Wonder scores 39.1, behind Astra (41.6) and Opusย 5.5 (41.2), so only the big sister takes first place.
Legal work
Harvey LAB-AA: Fullย 13.1, Wonderย 10.0. Astraย 8.6, Fableย 5.1ย 6.4, Opusย 5.5ย 4.2. A tough test where no model reaches even 20 yet.
Meta's internal overall index
Personal Superintelligence Index: Fullย 59.1. Wonder scores 54.3, just short of Opusย 5.5 (55.2) and Astra (54.5).
Even in Meta's own table, some subjects where the seniors still lead
Even big sister Full trails Kuroko's Opusย 5.5 in some subjects. Computer use (OSWorld V2.1): 78.4 vs 81.8. Building from scratch (ProgramBench): 59.0 vs 70.5. Terminal work (Terminal-Benchย 4.0-AA): 57.1 vs 59.6. Health questions (HealthBench Pro): 62.3 vs 65.6.
The model card adds its own caveats: competitor numbers include figures reported by each company, and Meta's in-house runs of other companies' models are "best-effort." MemoryBench and the Personal Superintelligence Index are Meta's own internal yardsticks.
A 14-month rebuild
Meta's own official blog describes Wonder as the result of a 14-month effort to rebuild its AI organization. The starting point was August 2025, when Meta reorganized its AI division into Meta Superintelligence Labs (MSL) under Alexandr Wang (TechCrunch). After Muse Spark in April 2026, Meta shipped something new almost every month from July: the image model Muse Image, Sparkย 1.1 through 1.3, Muse Code, and the personal agent Muse. According to Artificial Analysis, Wonder arrived about six weeks after Sparkย 1.3.
A deliberate delay of months
Zuckerberg said Meta delayed shipping by several months for safety and security (as quoted by Fortune and others).
Extra safeguards for chem and bio
Under Meta's own safety framework, it could not rule out the model reaching the "high risk" threshold in chemistry and biology without mitigations. So it added safety post-training and system-level filters that block dangerous inputs and outputs.
"Moderate or lower" after mitigations
With mitigations in place, Meta judged every domain "moderate or lower" and released it. The refusal rate on dangerous biology prompts (BioTIER, English) is 98.7%. Meta describes these as its strongest safety measures to date.
However, the detailed safety report has only been promised for the near future, and as of October 15, 2026 it has not been published.
Let's talk money
CNBC's headline framed the launch as Meta's latest test of whether users will pay. In other words: can Meta build a source of income beyond ads?
Aiming for a lower API price
Wang says the API is priced well below other companies' flagship models. At $4 / $20, it is less than half of Fableย 5.1 or Astra ($10 / $50), and the same as Kuroko's Opusย 5.5.
Cost per job
Artificial Analysis measured Wonder at $4.52 per task (about 2.8x Sparkย 1.3). Opusย 5.5 was $5.98 and Astraย $3.26.
The stock went up
On October 12 the stock closed up 3.38% ($742.96) โ on a day the Nasdaq fell (TIKR).
Some worrying signs
On October 9, Hunterbrook disclosed a short position and pointed out that weekly Muse downloads had fallen 8.1% (via TIKR).
September's tailwinds and headwinds
Muse, launched September 8, passed ChatGPT as the #1 free app on the App Store on September 21. Meanwhile Amazon blocked Muse, while Shopify welcomed it through Shop Pay (as reported in Academy News).
Next up: October 21
Third-quarter earnings. The first chance to check whether Wonder looks likely to turn into money.
The road to Muse Wonder
From the birth of MSL, through Meta-chan's enrollment, to earnings day.
Aug 2025: MSL is formed
The AI division becomes Meta Superintelligence Labs, led by Alexandr Wang.
Apr 8, 2026: Muse Spark
MSL's first model โ and the switch to closed models.
Aug 5: Muse Code
A coding tool for the terminal, announced alongside Sparkย 1.2.
Sep 2: Sparkย 1.3
Spark's third grade-up, about six weeks before Wonder.
Sep 8: personal agent Muse
Launched in the US โ a double that handles email, bookings and shopping for you.
Sep 21: #1 on the App Store
Muse passes ChatGPT as the #1 free app on the US App Store.
Oct 4: Meta-chan enrolls
Joins Hana AI Academy as a sociable new student who connects people and ideas.
Oct 12: Muse Wonder / Wonder Full
The first frontier-class sisters are announced (US time).
Oct 21: earnings (scheduled)
Third-quarter results โ the money check.
Report cards: Meta-chan vs. Kuroko
This time not Meta's self-reported numbers, but the independent Artificial Analysis overall report card (Intelligence Index v4.3, as of October 2026) โ
๐ Still top overall
- Overall Intelligence Index 58 (1st)
- Knowledge, AA-Omniscience: 46
- Even in Meta's own table, Kuroko leads in computer use and building from scratch
๐ Right behind, in 2nd
- Overall 56 (2nd). Younger sister Wonder: 52
- 1st in CritPt, a set of hard physics problems (by a narrow margin)
- Weak spot is knowledge: AA-Omniscience 30 (Kuroko: 46)
- Coding Agent Index 68 on Muse Code, tied with Claude Code + Sonnetย 5.5
The overall order: Kuroko's Opusย 5.5 (58) > Meta-chan's Wonder Full (56) > Gemi-chan's Geminiย 4 Argon, Chappy's GPT-6ย Astra and Kuroko's Fableย 5.1 (53 each) > Wonder (52). The previous-generation Sparkย 1.3 scored 48. Kuroko is still on top, with Wonder Full right behind her โ that's the report card as of October 2026. Note: the scores in our Astra feature (such as 66 for Fableย 5.1) come from a different version of the index, so they can't be compared directly.
The brand-new student has made it to just behind the front row.
Meta-chan's family spent 14 months rebuilding its whole AI organization, and has caught up to right behind Kuroko on the overall report card. Its strengths are subjects like "remembering" and "keeping going." Knowledge and building from scratch, though, are still the seniors' turf. And the real exam isn't the report card โ it's whether people will pay. The October 21 earnings and the rollout to the Muse app will be the next answer check.