Meta has put a frontier model on the table. On Monday, 12 October 2026, Meta Superintelligence Labs introduced Muse Wonder and Muse Wonder Full, which the company calls its most capable models for “personal and professional agentic tasks.” TechCrunch describes the pair as Meta’s first attempt at a frontier model, after the lighter, low-latency Muse Spark.
The pitch isn’t a better chatbot. Meta is going after agents that remember what matters to you and keep working on a goal until it’s done. Here’s what launched, who can use it, and why the reaction split almost right away.
What Meta Launched
There are two models. Meta says Muse Wonder is available now through the Meta Model API and Muse Code, its coding product. Muse Wonder Full, the bigger model, is available to Muse Code subscribers only. TechCrunch reports that Wonder Full won’t be served through the API at all.
Meta says Wonder will soon power its Muse assistant, starting with subscribers. It will also handle artifact generation and other tasks that need more horsepower. Later, Meta plans to use Wonder to build smaller, specialized models for everyone who uses Muse.
Meta says both models were trained at its gigawatt-scale Prometheus data center, which TechCrunch places in central Ohio. TechCrunch also notes this is the first frontier release since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August last year.
Active Memory and Active Focus
Meta has given names to the two traits it’s selling hardest.
- Active Memory: in Meta’s words, “carrying important context about you across conversations and actively retrieving the details relevant to the task at hand.”
- Active Focus: sticking with a hard goal by trying other approaches, working through blockers, and adapting to new information. Meta says the model asks for clarification when it needs it and checks in before important actions.
Meta’s demos lean into everyday agent work. In them, Wonder builds a playable 3D kitchen game from image inputs, plans a Saturday run in a browser and across apps, designs and orders custom swag based on Instagram photos, and works out a wedding seating chart with complicated rules. These are company demos, not independent tests.
“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 research scientist Kavya Kopparapu said in a launch video, per TechCrunch.
The Benchmark Claims (Meta’s Own)
Meta says the Wonder series “achieves state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding.” TechCrunch reports that in Meta’s own testing, presented in the model card, Wonder and Wonder Full score about the same as Fable 5.1 and GPT-6 Astra. On MemoryBench, Meta’s testing has Wonder ahead of Fable, Opus, and Astra.
Meta also lists the evaluations behind those claims. They include Harvey’s Legal Agent Benchmark for legal work, FocusBench for focus, and SWE-Marathon for long-running software projects. All of these numbers come from Meta’s own testing, and no independent lab had confirmed them at the time of writing.
Interestingly, Meta’s own post takes on the “benchmaxxing” criticism directly. It includes a chart meant to show that Wonder Full does better on tasks it wasn’t specifically trained for than its smaller predecessor’s trend would predict. Meta credits that to a larger, stronger base model.
Price and Access
According to TechCrunch, Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API. TechCrunch says that’s roughly the same as GPT-6 Sol or Opus 5.5. That price applies to Wonder only. If you want Wonder Full, the route is a Muse Code subscription, not the API.
That split matters when you read the launch charts. One commenter in the Hacker News thread pointed out that the headline scorecard figures map to Wonder Full, while the API price is for Wonder. So developers comparing API costs should test Wonder itself instead of assuming they’re getting the top numbers.
Safety, Per Meta
Meta says Wonder has its “strongest safeguards to date” against chemical, biological, radiological, nuclear, and cyber misuse, while still allowing defensive cyber work. It also calls Wonder its “most aligned model to date.” Meta says a full safety report with more evaluations is coming “soon.” Until that’s out, these are company claims.
What People Are Saying
The launch got loud fast. The official blog post hit the front page of Hacker News, and Techmeme’s lead item collected a long run of reactions on X. Early hands-on verdicts went both ways.
Shopify CEO Tobi Lütke was impressed (via Techmeme):
“This model is very very good. I was struggling to find anything to throw at it that it couldn’t do, even when other top models failed. Meta is back.” — Tobi Lütke (@tobi) on X
Hacker News user jjcm ran an image-to-HTML test and came away less convinced:
“I was expecting much more given the benchmark numbers it was putting out. I suspect its vision model isn’t mature yet, and that it’s been heavily optimized for text at this point in time.” — jjcm on Hacker News
In the same comment, jjcm said Wonder did “much worse than Opus 5.5” on that particular test. Both are personal takes from single users, not systematic evaluations.
JamoraquAI Take
Meta spent more than a year rebuilding its AI stack, and Muse Wonder is the first time it’s really competing with OpenAI and Anthropic at the top. The more interesting part is where it’s competing. Active Memory and Active Focus are a bet that the next fight is over agents that remember you and keep going through real work, not over who wins the next chatbot leaderboard.
Still, there’s a gap between the hype and the benchmarks, and it’s worth watching. The leaderboard wins come from Meta’s own testing. The strongest scores belong to Wonder Full, which most developers can’t call through the API. And early hands-on reports are mixed: one prominent CEO says he couldn’t stump it, while a Hacker News tester found its vision work behind Opus. Both can be true. Benchmarks measure what they measure, and your workload is the only test that counts.
Our advice hasn’t changed since the GPT-6 Astra release. Don’t switch your whole setup because of a launch-day chart. If you build agents, run Wonder on your own long, messy tasks and compare cost and reliability with what you use now. Treat “personal superintelligence” as a slogan until Meta’s safety report and some independent evals are out. Either way, Oct 12 settled one question: Meta is in the frontier race now.
Sources
- Meta — Muse Wonder: A Frontier Model For Personal & Professional AI Agents (12 October 2026)
- TechCrunch — Meta releases its first frontier model, Muse Wonder (Russell Brandom, 12 October 2026)
- Techmeme — 12 October 2026 archive (launch cluster and X reactions)
- Hacker News — Muse Wonder discussion
