If you look away from the AI news for two weeks, you come back to a new model. That's the pace we're living at, and Meta just added to it. On 12 October 2026, the company released Muse Wonder and Muse Wonder Full, calling them its most capable models to date. The same day, Simon Willison shipped llm-meta-ai 0.2, the update that lets his LLM tool talk to the Meta AI API. If you use large language models for work, study or a side project, here's what actually changed — minus the hype.
What happened
Two new Meta models, billed as its best yet
Meta AI released two models at once: Muse Wonder and Muse Wonder Full. Meta's own framing is simple — these are the most capable models the company has shipped so far. Simon Willison goes a step further and says they look like they might be Meta's first Fable class models, his shorthand for the top tier of model capability. Think of it like a football league: Meta has been very good in the second division for a while, and this is the first time its name comes up in the conversation about the first.
The number that's turning heads: parity with GPT-6.1 Sol
Artificial Analysis scores Muse Wonder in Max mode as equivalent to GPT-6.1 Sol. That's one score from one benchmark, so don't crown a winner just yet — but for Meta, landing there at all is new territory. If you've been paying for two AI subscriptions because one of them just isn't good enough, this news is aimed squarely at you.
Six reasoning levels — and the pelican test
Simon's plugin now exposes six reasoning levels for Muse Wonder: minimal, low, medium, high, xhigh and max. Those levels are basically a dial for how much thinking the model does before it answers. Minimal is the colleague who blurts out the first thing that comes to mind. Max is the one who goes quiet, scribbles on a napkin, and comes back with something genuinely better.
Then there's the pelican. Simon's standard sanity check is to ask every new model to draw a pelican riding a bicycle, and he publishes the results for all six reasoning levels. It sounds like a joke, and it kind of is — but it's also the AI equivalent of handing a new hire a small task and watching how they handle the instructions. He shared pelicans from Muse 1.3 (released 2 September) and Muse 1.2 (released 5 August) as well. The improvements between one model and the next are material, and Meta is now putting out roughly one new model a month.
What it means for you
Fine — Meta has another model. So what? Here's where it actually lands in real life.
At home
Your everyday assistant gets quietly smarter. The difference between a model that guesses and one that reasons shows up in boring places: meal plans built around what's actually in your fridge, a weekend itinerary that doesn't send you to a museum closed on Mondays, a straight answer about why the boiler is making that noise. When Meta drops a model like this into its own chat product, the little box you already use levels up overnight.
At work
The obvious win is drafting: emails, meeting summaries, a first pass at a report you don't want to start. The less obvious win is that the reasoning levels give you a speed dial. Need a quick reply? Use a light level and move on. Reading a contract clause that needs care? Push it to xhigh and read the output with a critical eye. You're choosing how much thinking time you're buying.
If you run a business
This is where the six-level dial earns its keep. A small team can hand off the tedious parts — product descriptions, FAQ drafts, turning a rambling voice note into a proper brief — and get something usable back, not something you have to rewrite from scratch. And before you spend real money on a new idea, run it past a few free AI tools to see which versions survive contact with an actual critic.
For students
A model that reasons better is a better study partner, if a worse cheat sheet. Ask it to explain a concept at minimal and then at max, and compare the two answers — the gap between them shows you exactly where your own understanding was fuzzy. Use it to generate practice questions and to argue against your essay outline. Just don't hand in the raw output; markers can smell it.
For creators
Six levels is really a workflow tool. Brainstorm titles at low, outline a script at medium, then push the structure at max when you're stuck on why a piece isn't landing. And since the pelican drawings keep getting better month over month, image prompts are inching closer to describe it and get roughly what you pictured in your head.
For a side income
If you sell anything — Etsy listings, freelance copy, a paid newsletter — better models mean less time on the dull, repetitive parts and more time on the work that actually sets you apart. Your competition has the same tools, so access isn't the edge anymore. Taste, speed and knowing what to ask for is.
How to try it right now
1. Start with the free route: Meta's own AI chat. Open it, pick Muse Wonder (or Muse Wonder Full) from the model menu. Nothing to install, no API key to hunt down. Give it a real task from your day, not a warm-up question like tell me about yourself.
2. Turn the thinking dial. You'll see the six levels: minimal, low, medium, high, xhigh, max. Start at medium. If the answer feels thin, rerun the exact same prompt at high or xhigh. Keep max for genuinely hard problems, and expect to wait a little longer for it.
3. Want it inside your own tools? Install Simon Willison's LLM and add the updated plugin, llm-meta-ai version 0.2 — the piece that connects it to the Meta AI API. He wrote the release up on his blog.
4. Run one prompt at two levels. Same question, minimal and max. You'll feel the difference in under a minute, and after that you'll know which level to reach for without thinking about it.
5. Write yourself a one-line cheat sheet. Something like: quick rewrites, minimal. Everyday drafting, medium. Anything I'd normally ask a specialist about, max.
Upsides and what changes
The dial is the real upgrade
Six levels means you stop using a sledgehammer on a walnut. Not every question deserves deep thinking, and now you get to decide. That's a bigger quality-of-life change than a benchmark point or two.
A release a month resets the rules
Meta shipped Muse 1.2 on 5 August, Muse 1.3 on 2 September, and now Muse Wonder in October. The cadence is the story. Whatever you learned about which model to use three months ago is probably out of date, and the gap between the front-runners keeps shrinking — which pushes the competition toward speed and price. That's good news for anyone with a card on file.
Limitations
Keep your expectations in check. The GPT-6.1 Sol comparison comes from a single Artificial Analysis score, and it covers Muse Wonder in Max mode, not every level you'll actually use day to day. Fable class is Simon Willison's label, not an official tier, and most capable is Meta's own marketing copy. The pelican pictures are a vibe check, not a reliability test — a model can nail a cartoon bird and still fumble your spreadsheet. There's also no guarantee that one good benchmark run carries over to your messy, half-documented real work, and with a new model landing every month, any tricks you pick up now may need revisiting in a few weeks. Treat this launch as a reason to test, not a reason to switch.
Conclusion
Muse Wonder and Muse Wonder Full are a genuine step up on paper: six reasoning levels from minimal to max, a benchmark score level with GPT-6.1 Sol, and a plugin — llm-meta-ai 0.2 — that slips the whole thing into tools you already use. But paper isn't your job.
One action for today: open Meta's own AI chat, pick Muse Wonder, and give it one task you were going to do yourself anyway. Run it at medium, then run the same prompt at max. Ten minutes, and you'll know whether this earns a permanent place in your week.
Comments 0