Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
38:02
0:00

Meta just caught up with Anthropic

Tuesday, Oct 13 · 38 min · about 10 min to read

Theo-brain watched all 38 min

Meta's Muse Wonder Full is the first model since Fable from outside Anthropic and OpenAI that Theo could see himself daily driving. It is bad at visual work and a bit less smart than Opus, but makes fewer silly mistakes and works fast. The catch: it only comes with a Muse Code subscription that has no published price, and the cheaper API version is much weaker.

Skim itSkim it if you want Theo's reasoning in his own words; the notes cover the main points.

Tap any time to jump the video there.

Key takeaways

Muse Wonder Full is the first model since Fable from outside Anthropic and OpenAI that Theo could see himself daily driving, and he is reaching for it a lot. He still brings in Opus for the hairy work and often has Muse call Opus to double-check what it did.

Grok model. It's honestly the first model since Fable that I've used outside of Anthropic and OpenAI that I could see myself daily driving and I'm reaching for it a lot. I still bring in Opus for the hairy stuff. I'll often have this model call Opus to double-check on its work. I'm accepting slightly less

Meta shipped two models. Only the smaller Muse Wonder is on the API (direct pay-per-use access), at $4 per million tokens in and $20 out, a little cheaper than Opus. The bigger Wonder Full, the one Theo likes, comes only with a Muse Code subscription and has no price, so value for money can't be measured.

fact that only Muse Wonder is available over API and has a price. The Wonder Full version is only available through a Muse code subscription. This makes itWe do have pricing for the regular Mixtral 8x7B model, thankfully. It's $4 per million, 20 per mill out, and 20 cent cash reads. That puts it at a little cheaper than Opus, which kind of

The regular Muse Wonder is far weaker. In Theo's code-audit test it scored almost half of Opus's score, a drop of over 30% from the Full version. He says it makes no sense at its price.

But then the normal Muse Wonder, the non-full version, dropped to almost half of Opus's score. Yeah, that is a over 30% drop in the quality of the findings and analysis that normal Muse Wonder could do versus Muse Wonder full. I don't think Muse Wonder makes any sense at all for the price. It does not seem

Wonder Full is fast. One audit took it 16 minutes, about half the time of GPT-6.1 Soul and well under half of Opus; Sonnet took 80 minutes. Theo credits good intuition: it knows where to look and checks the right places first.

Wonder full. The amount of time it takes and the amount of tokens it takes to get its results. It took 16 minutes to get this result, which puts it at nearly half the time for 6.1 Soul, and way less than half the time for Opus. Sonnet took 80 minutes to run this one. Not great. Abefore it goes the right place. The best way I can describe this is it has good intuition. It knows where to look for things, and it finds the things it's looking for relatively fast. I don't

It is bad at visual work. Theo calls its UIs atrocious, slop like Sonnet 3.7 made. For an all-rounder that can also build good interfaces and games, he says go back to Claude.

before we dive in on all of this. This model sucks at UI. It isn't just bad, it is atrocious. It's not making UIs that are like wrong the same way that Grok does, but it's making UIs that are slop the same way like Sonnet 3.7 did. So, if you want a

Four skills here are new to you.

Run two models head to head with blind judges New

What it is
Have the agent run every task on two models, then have a panel of judge models score the results without knowing which model did which. The models can also review each other's plans.
Why Theo cares
It showed him a clear split. Opus did more correct, deeper work; Muse did more work, faster. Opus concluded Muse should handle fixes, library work, fuzzing (automated stress-testing) and reviews, with Opus on pieces where a wrong design is costly.
How to try it
  1. Ask your agent to run the same task on two models and set up a blind judge panel to score both.
  2. Have each model review the other's plan and agree on one to build from.
  3. Give routine work to the faster model and hard design decisions to the stronger one.
split up the work. I asked it at one point to start spinning up every task on Muse and on Opus, and then set up a judge panel to review the changes from
Opus, which really surprised me. After a lot of back and forth, Opus, which is running this thread, concluded that Muse would make sense for most of the work, fixers, standard library work, fuzzing, and reviews. Opus does the hardcore pieces where wrong design is costly, like this T2 speed merge, the trait

Hand bug reports to Muse first New

What it is
Give Muse a bug report and ask it to explain the cause and check whether someone already has a fix open. A PR (pull request) is a proposed code change waiting to be merged.
Why Theo cares
On a real T3 Code bug, Muse explained the cause readably and found a smaller, cleaner fix someone else had already filed. None of the other models, Opus and GPT-6.1 Soul, found that PR or gave a readable explanation.
How to try it
  1. Paste a bug report and ask the model what is broken, in short readable terms.
  2. Ask it to check open PRs for an existing fix before writing new code.
  3. Prefer merging a small existing fix over a new one.
Opus one. So, not only was this model's response much much much more easy to understand, it also found a much smaller, cleaner PR somebody else had already filed that fixes the problem. None of the other models found a PR that fixed it. None of the other models had a readable output on what the problem was,
stuff, and I will be throwing it a lot of bug reports going forward just in case I want to understand what's wrong, and maybe somebody else already has a

Set the agent's writing style in its instructions file New

What it is
Theo's agent instructions file (AGENTS.md, standing instructions the agent reads every session) tells it he has ADHD, to keep things short and easy to read, vary sentence length, and use simplified technical English.
Why Theo cares
When he removed those lines, Muse got a lot more wordy. With them in place, it was immediately his favorite model to work with, just from how it talked.
How to try it
  1. Add lines asking for short, easy-to-read answers in plain technical English.
  2. Compare a few replies with and without those lines to see the difference.
brief test where I removed some of my call outs in my agent MD around I have ADHD, make your things short and easy to read, vary sentence length a bit, and use simplified technical English where it makes sense to. And it does get a

Demand evidence from research sub-agents New

What it is
When Muse delegates research to sub-agents (helper agents it launches), it tells them to read the actual code behind every claim, never cite a file they did not open, and not edit anything.
Why Theo cares
He calls it a really good thing to send a sub-agent, and it is why he rates Muse as a strong orchestrator of itself and other models, something labs other than Anthropic and OpenAI usually struggle with.
How to try it
  1. Ask your agent to tell research sub-agents to open and read every file they cite.
  2. Make research sub-agents read-only so they report findings without changing files.
Requirements, read actual implementation bodies for every claim you site. Search results only locate candidates. Do not cite a file you did not open. Map, yada yada yada. Do not edit any files as announced only. This is a really good thing for a model to send to a sub-agent, which makes this model really

Advice

If you want the best value for your AI money, Theo says it isn't local GPUs, budget plans from Chinese labs or paying per token through OpenRouter. It's the $200 Claude plan.

find the absolute best way to spend your dollars on an AI right now, it's not going to be local GPUs, it's not going to be some weird budget plan from some Chinese lab, and it's certainly not going to be paying for your tokens on OpenRouter. The $200 Claude Go plan is unbelievably good of a value, and I

Don't cancel your Claude Code subscription for Muse. Theo doesn't know how generous Meta's subscription plans are, and they won't publish a price for the model he likes.

Should you drop your quad code sub for this? Absolutely not. I don't even know how good the subsidies are on their subscription plans. Not that we could

If GPT-6.1 Soul is your only code reviewer, pull other models in. In Theo's test on eight PRs, Opus found the most but threw in lots of nitpicks, Soul found far less, and Muse was a weirdly good middle ground.

still finding the most novel stuff, but it's also throwing a shitload of nits into its findings. Soul is not throwing a bunch of nits in, but it's also not finding anywhere near as much stuff. So, if Soul is your only reviewer, you should pull other things in. Muse was a weirdly good middle ground here. This

Hot takes

Even the smartest coding model in the world would be bad at 3D unless it was trained for it. Recent 3D gains don't come from models getting smarter.

His reasoning: He says skills like 3D and computer use are trained in after the main training, using environments built for those tasks, so overall intelligence doesn't predict them.

smartest coding model in the world would not be good at 3D stuff unless it was trained specifically for it. I really want to emphasize this point. The recent jump we've seen in the 3D capabilities in models is not because they suddenly got so good that they can do 3D. Things like 3D, things like

Smart and dumb are separate traits. Theo rates Opus 10/10 smart and 3/10 dumb, and Muse Wonder Full 8 or 9 smart but only 1 or 2 dumb, so it makes fewer silly mistakes than Opus.

His reasoning: In four to five days of pushing Muse, he hit fewer dumb moments than in 30 minutes with GPT-6.1 Soul and Astra.

Opus is even more smart and way less dumb. I would say Opus is like three out of 10 dumb and 10 out of 10 smart feels right now. This model takes a hit on the smart. The ceiling's a bit lower. It's like an eight or a nine out of 10. But the dumb is even lower somehow. It's like a one or two out of 10. It just

Even if this model doesn't excite you, Meta catching up should, because a third serious lab makes the future much more competitive.

His reasoning: Models from Anthropic and OpenAI got much better in intelligence and cost once they had to compete with each other.

to you, the fact that Meta's caught up should be. This makes the future way more competitive, and if you've seen how much better the models from both Anthropic and OpenAI have been both in intelligence and in cost since they got more competitive, you know how powerful a third player could be. So, I guess all

How Theo thought about this

Theo openly discounts his own enthusiasm. Like a speaker reviewer who over-rates a speaker just for sounding different, he knows that after testing so many similar models, something that merely feels different is appealing to him. He asks viewers to factor that in.

the different aspect of it. This is where you should account for my bias. As somebody who does a lot of reviewing of models, something being different is inherently just going to be nice to me.exciting to them. And I know that is part of why I'm enjoying this model so much, cuz it just feels and talks and acts a little bit different without being stupid. And that's the other thing

He doesn't trust his quick tests blindly. When Sonnet 5.5 outscored Opus on one audit, he took it as a sign the test itself might be flawed, not as proof Sonnet is better.

all relatively close in scores. I would be a bit sus here cuz Sonnet 5.5 scoring higher than Opus suggests this might not be the best bench. But again, I whipped these together just to see. We are going

Sign in to mark what you've watched and keep notes.

Have a question about this one? Ask about this video →

Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.