The most expensive Pokémon run i've ever seen cost $96.06 and never left the first forest.
Somebody ran Muse Wonder, Meta's new frontier model, at Pokémon Red on launch day. It picked its starter by turn 20 and reached Viridian City by turn 59, right on pace with the fastest agents. Then it walked into Viridian Forest on turn 208 and spent the next 792 turns in there.
Six hours of stream. Almost a hundred dollars. It never beat Brock.
Opus 5.5 played the same game at the same token price and won in 271 turns. It cost $8.44.
Here's why that matters. The launch week was all triumph. Meta called the model a major leap toward personal superintelligence, and Wired crowned the company back at the frontier.
Hacker News gave it the highest compliment it knows: i'm sorry zuck, i wasn't familiar with your game.
And right under that headline, the flagship was paying $12 a minute to wander through a forest.
The gap between the poster and the product is the real story.
Why people lost their minds
i used to think Meta would never catch up in AI. The labs with the best models were the ones that bet everything on them, and Meta always felt like it was running a side project with an ad business attached.
Then the numbers came out.
Muse Wonder launched Oct 12, and Wonder Full with it. Wonder runs through the Meta API at $4 per million input tokens and $20 per million output.
Wonder Full, the stronger one, is locked inside Muse Code subscriptions at $50 or $200 a month. Consumers get it through the Muse app at $20 a month, if they feel like paying.
| Where | Price | What you get |
|---|---|---|
| Muse app | $20/mo | Wonder |
| Muse Code | $50 or $200/mo | Wonder + Wonder Full |
| Meta API | $4/$20 per MTok | Wonder |
The headline features are Active Memory and Active Focus. Meta says the model remembers what matters to you across conversations and keeps pushing toward a goal instead of stalling. What that means in practice:
- try a different approach when one fails
- keep working through blockers
- check in before doing something important
On MemoryBench it beats Fable, Opus, and Astra. On the Artificial Analysis index, Wonder Full sits second, behind only Opus 5.5.
Is the scoreboard real
here's a question people keep asking me: is the scoreboard real?
There are literal bets on it. People are gambling on Polymarket and Kalshi over how much of Wonder's lead comes from training on the benchmarks themselves. The top comment on the launch thread asked the same thing, straight up: how much benchmaxxing has been done here?
Meta saw this coming. Its own launch post has a section defending against the accusation, and it coins a phrase for the opposite quality: big model smell, the feeling that a model handles stuff it was never trained on way better than the numbers predict.
look, i get why people are excited. But scores are a starting point, not a review. Every frontier lab trains toward the benchmarks it publishes.
That's not fraud, it's how the game is played. And Wonder's real review is happening in the forest right now.
The comeback nobody voted for
what actually happens is a redemption arc nobody voted for.
A year ago this lab was a punchline. Zuckerberg bought Scale AI for billions, put a twenty-something Alexandr Wang in charge of Meta Superintelligence, and pushed out Yann LeCun, a Turing Award winner, over his loud complaints about LLMs.
Insiders laughed. Wall Street hated the capex. The Metaverse had already burned through money like it was nothing.
Then Muse Spark shipped and quietly became a workhorse. People on Hacker News said it moved 90% of their coding work off Codex. And now Wonder Full lands second on the index, one spot under Opus 5.5, and the same people are acting like they always believed.
i'm sorry zuck, i wasn't familiar with your game
I'm not going to pretend I saw it coming. I called Meta's AI push a vanity project last year. I was wrong about the model.
I still think the company has a trust problem, and we'll get to that.
The race gets cheaper
the problem isn't what you think. It's not the benchmark scores.
If Meta can buy and scale its way to second place, then frontier models stop being a moat. One comment on Hacker News put it plainly: give any country with 20 to 40 billion dollars a couple of years and it will have its own champion model. That's what commoditization looks like, and it's already happening.
Meta also has the distribution to make it sting. WhatsApp holds three times the users of ChatGPT's entire app.
If Wonder quietly powers a good enough assistant inside apps people already open, nobody needs to download a thing. That's the part OpenAI can't copy by making a smarter model.
And OpenAI is having a rough week. Its pro subscription value just got halved, and the IPO talk is wobbling.
The story here isn't only about Meta winning. It's about the whole market getting cheaper and faster at once.
Watching the machines play
Separately, i've been watching people turn AI game runs into a spectator sport. PokéBench keeps a leaderboard of agents playing Pokémon Red, with turn logs and cost per run. It's like horse racing, except the horses are language models and the track is a 1996 game.
It takes me back to watching my cousin play on a Game Boy in the back of a car. Same sport, way higher stakes.
He'd get stuck in Viridian Forest too, then hand me the device and tell me to fix it. I never could either. We'd restart the whole save and pretend it didn't happen.
The difference now is that the AI doesn't ask for help when it's stuck. It just keeps buying more tokens and walking deeper into the trees.
That's the most human thing about it, honestly. i've paid for dumber persistence.
Who should skip this
Most people don't need Muse Wonder. That's not a hot take, that's the price tag talking.
The free Muse app already handles the daily stuff: reminders, notes, small tasks. The moment you jump to Wonder you're paying $20 a month for the memory features, and $100 for the heavier tier. The people who'd actually benefit, the ones running long autonomous jobs, are already on Muse Code at $50 or $200.
And let's talk about the privacy stuff, because it's the part everyone is dancing around. Users complained the Muse agent read private messages without permission.
Meta disputed it, and the AI chief said users may have forgotten what they authorized earlier in a conversation. Maybe that's true.
Maybe it's the creepiest sentence of the year.
Both can be true at once. If you don't trust Meta with your inbox, Wonder doesn't fix that. It makes it sharper.
The forest is still there
i keep coming back to that $96 walk. Not because it proves Wonder is bad. It proves what frontier means right now: a model that can do a hundred useful things and still, with total confidence, spend six hours going nowhere.
They'll fix it. The next run will clear the forest and the headline will change.
But i'll remember the first one. The expensive, honest, slightly embarrassing first one. And next time somebody tells me a model is superintelligent, i'm going to ask one question: did it get past the forest yet?