Muse Wonder Gets Lost in the Forest: A Frontier Model's $96 Faceplant

Someone paid $96.06 to have Meta's new frontier model, Muse Wonder, play Pokémon Red. It couldn't even make it out of the first forest.

According to analysis from a Bilibili creator and a YouTube channel covering the livestream, Muse Wonder went live with the game on launch day. It picked its starter on turn 20 and reached Viridian City by turn 59, keeping pace with the fastest agents. Then, on turn 208, it walked into Viridian Forest — and spent the next 792 turns stuck inside. Six hours of livestreaming, nearly a hundred dollars, and it never even beat Brock, the first gym leader.

Same game, same token pricing: Opus 5.5 finished in 271 turns for $8.44.

This is worth talking about because launch week was nothing but triumph. Meta called the model a major leap toward personal superintelligence. Wired handed the company back its frontier crown. Hacker News offered the highest praise it can give: "i'm sorry zuck, i wasn't familiar with your game."

And right beneath that headline, the flagship model was lost in a forest at $12 a minute. The gap between the poster and the product — that's the real story.

Pricing and Capability Layout

Muse Wonder launched on October 12, alongside Wonder Full. Wonder is available through the Meta API at $4 per million input tokens and $20 per million output tokens.

The more powerful Wonder Full is locked behind a Muse Code subscription at $50 or $200 per month. Regular consumers can access it through the Muse app for $20 a month — if they're willing to pay.

Channel	Price	What You Get
Muse App	$20/month	Wonder
Muse Code	$50 or $200/month	Wonder + Wonder Full
Meta API	$4/$20 per million tokens	Wonder

The headline features are Active Memory and Active Focus. Meta claims the model can remember what matters to you across conversations and keep pushing toward goals rather than stalling. In practice, that means: trying a different approach when one fails, pushing forward when blocked, and confirming before taking important actions.

On MemoryBench, Wonder beats Fable, Opus, and Astra. On the Artificial Analysis index, Wonder Full ranks second, just behind Opus 5.5.

Active Memory and Long-Horizon Autonomous Tasks: Active Memory refers to a model persisting user preferences, project context, and past decisions across sessions, and proactively recalling them in later conversations instead of relying on the user to repeat themselves. It's typically implemented with an external vector store or structured memory layer, paired with retrieval and summarization compression to control context length. Active Focus, meanwhile, is the mechanism for maintaining multi-step execution toward a set goal — switching strategies after failure, pushing through obstacles — essentially a combination of long-range planning and tool use. These capabilities are extremely sensitive to token consumption: long-horizon tasks repeatedly read and write memory and retry failed steps, so under output-token billing, a single multi-hour autonomous run can easily burn tens of dollars. That explains why Wonder's Pokémon run cost so much more than a short conversation.

Can You Trust the Benchmarks?

People are betting real money on this question. On Polymarket and Kalshi, they're wagering on how much of Wonder's lead comes from training directly on benchmarks. The top-voted comment on the launch post cuts straight to it: how much benchmaxxing went on here?

Meta saw it coming. The official launch post includes a dedicated section rebutting the accusation, and even coins an antonym — "big model smell" — to describe that feeling when a model performs far beyond what the numbers predict on tasks it was never trained on.

It's easy to understand the excitement. But benchmarks are a starting point, not an evaluation. Every frontier lab trains toward its own published benchmarks. That's not fraud — that's the rules of the game. And Wonder's real evaluation is happening right now, in the forest.

Benchmaxxing and Benchmark Contamination: Benchmaxxing refers to labs — intentionally or not — tuning training data and post-training objectives toward public benchmarks (MMLU, GPQA, SWE-bench, etc.) so that scores on those benchmarks are inflated without reflecting true generalization. Common tactics include mixing benchmark questions or their variants into pretraining corpora, using benchmark scores as reward signals during RLHF, and only disclosing favorable benchmarks. Since major labs all publish self-selected benchmarks, cross-comparisons are often distorted — which is why the community demands third-party independent evaluation and held-out sets. Real-world evaluations like PokéBench, with turn logs and per-step cost records, draw attention precisely because they're hard to cover with training data directly, exposing genuine weaknesses in long-horizon planning and self-correction.

The Comeback Nobody Voted For

What actually happened is a redemption narrative nobody voted for.

A year ago, this lab was a punchline. Zuckerberg spent billions acquiring Scale AI, put twenty-something Alexandr Wang in charge of Meta Superintelligence, and pushed out Yann LeCun over his fierce criticism of LLMs. Insiders laughed. Wall Street hated the capex. The metaverse was already burning cash by the truckload.

Then Muse Spark launched and quietly became a workhorse. Someone on Hacker News said it moved 90% of their coding work over from Codex. Now Wonder Full ranks second on the index, just behind Opus 5.5, and the same crowd acts like they were bullish all along.

I'll admit it: last year I called Meta's AI push a vanity project. On models, I was wrong. But I still think this company has a trust problem, which I'll get to.

The Race Gets Cheaper

The problem isn't the benchmark scores.

If Meta can buy and scale its way to second place, the frontier model is no longer a moat. One Hacker News comment put it bluntly: give any country $20–40 billion and a few years, and it'll have its own champion model. That's what commoditization looks like, and it's already happening.

Meta also has a distribution advantage. WhatsApp has three times the user base of the entire ChatGPT app. If Wonder quietly powers a good-enough assistant inside apps people already have open, nobody needs to download anything. That's something OpenAI can't replicate with a smarter model.

OpenAI had a rough week too — the Pro subscription's value was just cut in half, and IPO discussions are wavering. The story isn't just that Meta won. It's that the whole market is getting cheaper and faster at the same time.

Model Commoditization and Distribution Moats: Model commoditization means frontier capabilities are replicated by multiple labs at similar cost and within similar timeframes, so "having the strongest model" no longer constitutes a durable competitive advantage — profit shifts toward distribution channels, data entry points, and ecosystem migration. Cloud computing and search went through similar paths historically: once the technology lead window narrows, the winner is often whoever controls the user's default entry point. Meta's WhatsApp and Instagram have billions of monthly active users; if an assistant is embedded by default in those apps, users get "good enough" intelligence without migrating, which erodes subscription models built around model quality. For pure model vendors like OpenAI, the moat is forced to shift toward brand, developer ecosystem, and enterprise integration, while the price and capability gap between models keeps compressing.

Watching Machines Play Games

Separately, I've been watching people turn AI game runs into a spectator sport. PokéBench maintains a leaderboard of agents playing Pokémon Red, complete with turn logs and per-run costs. It's like horse racing, except the horses are language models and the track is a 1996 game.

It reminds me of watching my cousin play Game Boy in the back seat as a kid. Same sport, much higher stakes. He'd get stuck in Viridian Forest too, then hand me the device to figure it out. I couldn't either — we'd restart the save and pretend it never happened.

The difference is that when an AI gets stuck, it doesn't ask for help. It just keeps buying tokens and walking deeper into the woods. Honestly, that's the most human thing about it. I've paid for stupider persistence.

PokéBench and Agent Game Evaluation: PokéBench is a class of public leaderboards that use classic games like Pokémon Red to evaluate the long-horizon capabilities of language model agents, recording completion turns, per-step decision logs, and per-run costs. Games provide fixed rules, clear goals, and sparse rewards, testing planning, memory, exploration, and recovery from failure simultaneously. And because game footage and walkthroughs are widely available online, there's an inherent risk of training data contamination — which is why these leaderboards typically emphasize cost and turn trajectories rather than a single win rate. Compared to static Q&A benchmarks, this kind of "watchable evaluation" exposes model weaknesses on livestream, making it an important complement to the community's efforts to verify whether benchmark scores reflect reality.

Who Should Skip This

Most people don't need Muse Wonder. That's not a hot take — it's the price tag talking.

The free Muse app already handles everyday tasks: reminders, notes, small jobs. Once you jump to Wonder, you're paying $20 a month for the memory features, or $100 for the heavy tier. The people who actually benefit — those running long-horizon autonomous tasks — are already on the $50 or $200 Muse Code plan.

And the privacy issue has to be mentioned, because it's the part everyone tiptoes around. Users have complained that the Muse agent reads private messages without permission. Meta denies it, and the AI chief says users may have forgotten what they authorized earlier in the conversation. Maybe that's true. Maybe it's the most chilling sentence of the year. Both can be true at once. If you don't trust Meta near your inbox, Wonder won't fix that — it'll just make it sharper.

The Forest Is Still There

I keep coming back to that $96 trek. Not because it proves Wonder is bad, but because it proves what "frontier" means right now: a model that can do a hundred useful things while confidently spending six hours going in circles.

They'll fix it. The next run will make it out of the forest, and the headline will change. But I'll remember the first one. That expensive, honest, slightly embarrassing first one. Next time someone tells me a model is superintelligent, I'll ask one question: did it make it out of the forest?