Key takeaways
- Meta's own testing, presented in a model card, puts Muse Wonder at scores comparable to Claude Fable 5.1 and GPT-6 Astra.
- Artificial Analysis, as reported by Mashable, placed Muse Wonder Full behind only Claude Opus 5.5 Max on its Intelligence Index, level with Claude Sonnet 5.5 Max.
- Wonder Full is limited to Muse Code subscribers and isn't served through the API, so the model with the best independent ranking is not the one developers can call directly.
- How Wonder compares with Google's Gemini 4 is contested: Artificial Analysis shows Wonder ahead, while The Verge says the picture against Gemini 4 Argon is unclear.
On Monday, October 12, 2026, Meta launched a pair of AI models, Muse Wonder and a larger variant called Muse Wonder Full, and says they put the company at the frontier of AI. An independent benchmarker, Artificial Analysis, mostly backs that up. But the two scorecards don’t line up neatly, and the differences matter if you’re choosing a model to build on.
What exactly did Meta release?
Two models from Meta Superintelligence Labs. Muse Wonder can be reached via Meta’s API and in Muse Code. Muse Wonder Full is limited to Muse Code subscribers and isn’t served through the API, according to The Verge, TechCrunch, Mashable and CNBC.
They follow Muse Spark, which debuted in April 2026 as a lightweight, low-latency model. TechCrunch says the Wonder models are larger and more capable, and score higher on benchmarks and on complex work like coding. Meta says they will also power its Muse agent, with paid subscribers getting access first or only. Coverage disagrees on exactly which model and when, and Meta has given no timeline.
Labels like “first frontier model” (TechCrunch) and “back at the frontier” (Wired) are those outlets’ characterizations. Treat them as framing, not as measurements.
What does Meta claim about its own models?
Meta’s model card puts the models at benchmark scores comparable to Claude Fable 5.1 and GPT-6 Astra. On Artificial Analysis’s Intelligence Index, Meta says its largest model beats Fable 5.1 and lands just under Claude Opus 5.5. These are Meta’s numbers, from Meta’s testing.
Wired reports that on certain specialized tests, such as legal-document work and software building, Meta claims wins over GPT-6 Astra, Fable 5.1 and Opus 5.5. TechCrunch adds that Meta says Wonder scores higher than Fable, Opus and Astra on MemoryBench, a test of continual learning and memory.
Memory is Meta’s headline feature. Research scientist Kavya Kopparapu told TechCrunch: “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.” Meta pairs it with Active Focus, which it describes as staying on a goal, trying other approaches, recovering from errors and checking in before important actions.
The individual benchmark scores behind these claims aren’t published in the coverage. You get rankings and comparisons, not a table of results.
The independent ranking
Artificial Analysis placed Muse Wonder Full behind only Claude Opus 5.5 Max on its Intelligence Index, level with Claude Sonnet 5.5 Max, according to Mashable. Mashable also reports that Artificial Analysis shows Wonder ahead of Gemini 4, though The Verge calls that comparison unsettled.
The same report says Muse Wonder Full (max) lands in the top three on six of ten Index evaluations, and stands out on scientific reasoning and agentic knowledge-work tasks.
Two cautions. Mashable is the only outlet in the coverage that details this ranking, so there’s no second confirmation. And “independent” is Mashable’s framing. Artificial Analysis does run its own index, but Meta also cites that index in its own pitch, which makes it a shared yardstick, not a neutral referee. The ranking also covers the max setting of Wonder Full, not every configuration.
Where the scorecards disagree
The sources don’t tell one story. They differ on Anthropic, Google, price and training, and it helps to see each point side by side.
- Versus Anthropic: Meta says its largest model sits above Fable 5.1 and just below Opus 5.5. Artificial Analysis, via Mashable, has Wonder Full behind Opus 5.5 Max and level with Sonnet 5.5 Max. Both put Opus 5.5 on top. They name different rivals below it.
- Versus OpenAI: Meta’s model card says scores are comparable to GPT-6 Astra, and Meta claims wins on some specialized tests. Where Artificial Analysis places Astra isn’t reported.
- Versus Google: Artificial Analysis shows Wonder ahead of Gemini 4. The Verge says it’s unclear how the models compare with Gemini 4 Argon, which is still being sized up.
- Price: TechCrunch reports API pricing of $4 per million tokens for input and $20 per million for output, which it says is roughly the same as GPT-6 Sol or Claude Opus 5.5. Meta’s Alexandr Wang told CNBC the API will cost more than standard subscriptions but be priced “meaningfully below” leading OpenAI and Anthropic models.
- Training: TechCrunch says Meta’s post credits its Prometheus data center in central Ohio. Wired says Meta didn’t disclose details of how the models were built or trained.
Why this matters if you’re choosing a model
If you build on APIs, the key detail is access. The model with the best independent ranking, Wonder Full, isn’t available through the API. You can call Muse Wonder, but the coverage doesn’t say how that model scores on Artificial Analysis’s Index.
So I wouldn’t read the ranking as a sign that the API model is second best. I’d read it as evidence that Meta’s top configuration is competitive, and then test the one you can actually buy.
On cost, Wonder Full requires a Muse Code plan, which CNBC reports at $50 or $200 per month. The Muse agent plans that include Wonder run $20 or $100 per month. For the API, the price gap against OpenAI and Anthropic is disputed, as above, so run your own token math before you migrate anything.
The most useful benchmark is your own workload. Meta’s claimed wins in legal-document work and software building are Meta’s claims, and Wired itself cautions that benchmarks aren’t always reliable. Build a small test set from your real tasks and score every model on it. Chip Huyen’s book AI Engineering is a well-regarded guide to evaluating foundation models if you haven’t set that up before.
Safety claims deserve the same scrutiny. Meta says the model has its strongest safeguards yet against cyber misuse, and that for now these come only with the paid model. A full report on its safety evaluations is promised as coming soon. Some users have also claimed on social media that the Muse agent accessed personal messages without authorization. Meta disputes that, and Wang has suggested some users authorized access earlier in a conversation and forgot. If you’re weighing agents across vendors, our look at Google Gemini agent email identities covers the identity questions IT teams are asking. For how safety disputes have played out at a rival lab, there’s our timeline of OpenAI’s firing of safety researchers. And if your team already leans on OpenAI’s chat product, what ChatGPT Intelligent UI changes is a useful baseline.
What’s still unknown
Plenty. The biggest gap is a second independent check. Right now one outlet reports the Artificial Analysis ranking in detail, and the per-benchmark numbers aren’t in the coverage.
Other open questions:
- Whether Wonder Full will ever reach the API or non-Muse Code users.
- How Wonder compares with Gemini 4 Argon.
- Which model will power the Muse agent, and when.
- What the full safety evaluations show.
- How the models were trained, given the conflict between TechCrunch and Wired.
- Whether larger open-weight models arrive. Meta has released a small one, Muse Glimmer, and says bigger ones are planned.
Bloomberg also covered the launch, but only its headline was available: Meta says the new Muse model reaches the AI frontier after an aggressive push. Beyond that, no details can be attributed to it. Whether the benchmark gains hold up in everyday use is something only developers running their own tests will find out.
Tools & products mentioned
- AI Engineering on Amazon — Check current price
Disclosure: some links in this article are affiliate links or our editor's own courses and book. If you buy through them, TechyCluster or its editor may earn money at no extra cost to you. This does not influence what we cover or how we cover it. As an Amazon Associate, TechyCluster earns from qualifying purchases.
Frequently asked questions
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder can be used via Meta's API and inside Muse Code. Muse Wonder Full is limited to Muse Code subscribers and isn't served through the API.
Is Muse Wonder better than Claude Opus 5.5?
Not according to the rankings reported so far. Artificial Analysis, via Mashable, placed Muse Wonder Full behind only Claude Opus 5.5 Max. Meta also puts its largest model just under Opus 5.5 on the same index.
How much does Muse Wonder cost?
TechCrunch reports API pricing of $4 per million tokens for input and $20 per million for output. CNBC reports Muse agent plans with Wonder at $20 or $100 per month, and Muse Code plans at $50 or $200 per month.
Does Muse Wonder beat Gemini 4?
Artificial Analysis shows Muse Wonder ahead of Gemini 4, per Mashable. The Verge says it's unclear how the models compare with Gemini 4 Argon, so the answer depends on which Google model you mean.
Get the next one in your inbox. One email a day with the tech stories that matter, explained in plain English. Subscribe free.
Sources
This article was compiled from reporting by the following outlets. Links go to the original reports.
- The Verge: Meta has some Wonder Full new AI models
- TechCrunch: Meta releases its first frontier model, Muse Wonder
- Mashable: Meta launches Muse Wonder, catching OpenAI and Anthropic at the frontier
- CNBC Tech: Meta debuts Muse Wonder AI model in latest test of users' willingness to pay
- Wired: Muse Wonder Puts Meta Back at the AI Frontier
- Bloomberg.com: Meta Says New Muse Model Reaches AI Frontier After Aggressive Push
