Muse Wonder is Meta's first frontier model, and you can't download it

Mark Zuckerberg posted a simple card on October 12: Muse Wonder and Muse Wonder Full are here, and the bar chart says Meta's new model family beats Claude Opus 5.5 and GPT-6 Astra on the Personal Superintelligence Index. Four million views later, the fight over what that actually means has settled into three camps. Meta's numbers say frontier. The independent index says nearly-frontier, second place. And a regional podcaster named Jacob, not to be confused with the Albion Bay one, put it most plainly on Hacker News: it is "an intelligent, almost SOTA, but way too expensive" model.

All three are right, and the interesting story is not the scoreboard. When Meta's chief AI officer Alexandr Wang sat down with John Collison on Cheeky Pint two days after launch, the most revealing thing he said had nothing to do with benchmarks. It was about how these models get good at things: reinforcement learning runs on environments somebody deliberately builds, with a reward somebody deliberately defines. So the question "how good is Muse Wonder?" has a sharper version: good at what, exactly, and why that list is shorter than it looks. Muse is genuinely excellent at coding. It is not yet good at making video games. That is not a talent gap. It is a training-data gap, and it tells you exactly how to read every frontier launch from here on.

What Meta actually launched

Two models, one family, both closing at a 1M-token context window. Both take text, images, video and audio as input and reply in text only. The identities were confirmed against Meta's model card and dev.meta.ai pricing page:

That second row is the part people keep missing. Wonder Full, the model Meta put on the front page of the launch card, has no API and no per-token price. You cannot buy it. The only way to use it is to subscribe to Muse Code, Meta's coding agent, at $50 or $200 a month. Reviewers therefore cannot price it, developers cannot ship with it, and every "beats Opus" chart you saw this week is advertising a product you can only rent inside Meta's own app. One Hacker News commenter summarised the strange distribution: Wonder "will soon power experiences in" the assistant app, while Full is meant for API usage and "Muse Code" only, "It's getting confusing what having 'Muse' means."

Two renewals ago Meta was the company of Llama, the weights-everywhere bet that pushed the whole industry open. Muse Spark started closing that door in April 2026 when it shipped with no weights at all. In August Meta released Muse Glimmer, a 30B Apache 2.0 model, and Zuckerberg promised Spark 1.2 weights "coming soon". Five weeks later that promise is still unfulfilled, and the frontier family is fully closed. You can draw the strategy shift in one line: the man whose open-weights era commoditised everyone else's models is now charging Opus prices for his own.

How good is it, really?

Start with the numbers, because this model is unusually well documented. Meta's Personal Superintelligence Index is an internal composite of seven personal and knowledge-work evaluations. On it, Wonder Full (max) scores 59.1, ahead of Claude Opus 5.5 at 55.2, GPT-6 Astra at 54.5, Claude Fable 5.1 at 49.9, and plain Wonder at 54.3. That is the 59.1 you saw everywhere. The catch is in the footnote: the composite is Meta's own, the evaluation mix is Meta's own choice, and the referee wrote the exam.

So step outside. Artificial Analysis runs an independent index, v4.3, that it computes itself from ten evaluations nobody at Meta touched. There the order flips: Opus 5.5 leads at 58, Wonder Full scores 56, and its little sibling Wonder scores 52, just ahead of GPT-6 Astra and Gemini 4 Argon at 53. Both models beat Spark 1.3's 48, which is the real proof of the six-week sprint. Second place at the frontier, six weeks after your previous release, is genuinely impressive. It is just not the "state of the art for personal superintelligence" the launch card implies.

A few rows of the card are worth pausing on. Terminal-Bench 4.0 is the biggest jump: Spark 1.3 scored 33, Wonder Full scores 57.4, which lands it level with Gemini 4 Argon in Artificial Analysis's table. MemoryBench 85.5 is the best score of any model tested, and that matters: memory is, as Wang put it on the podcast, "one of the clear requirements to build reliable agents" and an area Meta invested in "disproportionately". HealthBench Pro is the one clear loss, with Opus 5.5 and Astra well ahead, and on AA-Omniscience, which tests whether a model actually knows things rather than guessing pretty, Wonder Full scores 30 against Opus 5.5 at 46. Hallucination rate, though, is excellent: 23 percent, second only to Gemini 4 Argon's 15 percent.

Add the model card's own candour and the picture sharpens. The card admits its SWE-Marathon row is not an apples-to-apples comparison, because the rival scores were recorded inside Codex and Claude Code harnesses while Meta's was reported by a third party at a different reasoning effort. The card also discloses that 71 of 92 ExploitGym flags the model captured were found through unintended vulnerabilities; only 21 count by the ranking's own rules. Publishing both of those is to Meta's credit, and they still shave the headline numbers made of the same metal.

Excellent at coding, not yet at making games

Now the part that got buried under the benchmark fight, and the part that actually matters if you buy these models.

Coding, Muse Wonder Full is uncomfortably good. On SWE-Marathon, an ultra-long-horizon software engineering benchmark graded by executing real project tests, it scores 57.5 percent against Opus 5.5's 46.3. On DeepSWE 1.1 it is 79.0 against roughly 74 for Opus and Astra. Terminal-Bench jumped from 33 to 57 across one model generation. Vals AI's independent index had the model debuting ahead of every OpenAI model, up 6.4 points from its predecessor. In the launch demo Wang showed it building a playable 3D game from a few images of a person's desk. The agentic loops work: the model grinds, retries, switches Approach, and asks before doing something irreversible. That is Active Focus doing exactly what it was trained to do.

But then try it on the thing the demo hinted at, actually designing and finishing a video game, and the frontier sheen drops fast. It can code a game when you spec the mechanics. Nobody is getting from Muse Wonder an original, balanced, finished game that is fun to play, the way they were getting playable shockers from Anthropic and OpenAI models months ago in one-shot 3D and interactive demos.

And here is why, straight from Wang's mouth: models get wonderful at things through reinforcement learning, and RL only works on tasks somebody has turned into environments with a reward. On the podcast he was asked how post-training works. "First is just producing the environments and producing the tasks", he said, "and I think that is one of the areas where there's a lot of intentionality in training these models." Then came the detail that explains everything about what models are good at: "In a lot of cases, less is actually more when it comes to data in RL." A few great environments beat a mountain of mediocre ones.

So follow that logic. Where does Muse Wonder's training compute go? Into environments where success is checkable: repos whose tests either pass or fail, terminals whose commands either work or do not, financial tasks with right answers, legal deliverables a judge panel can score against criteria. Nobody at Meta, or anywhere, has built the RL environment whose reward function is "is this game fun?" There is no test suite for joy. So the model never got RL'd on game design, and the skill is simply not there in the way coding skill is. Zuckerberg said as much himself, describing the pipeline: SFT bakes a behaviour in at 10 percent of the time, and "in reinforcement learning, you really sort of give the model reinforcement in various situations to bring the elicitation of those behaviors dramatically up." No RL on game-fun, no dramatic uplift on game-fun.

This is, quietly, the most useful idea in the whole launch, and it generalises: a model will be superhuman exactly where somebody paid to build reward environments, and mediocre everywhere nobody has. Coding has enormous paid environments. Game design, taste, judgement in ambiguous domains: not yet.

One Hacker News thread on the launch turned into a six-hundred-post debate about whether Meta "benchmaxxed" the model, training to the test. The honest answer is subtler and more interesting. Every lab trains to some test. The question worth asking of any launch is now: which tasks got real RL environments, and which are riding on pretraining alone?

The pricing is the message

Muse Wonder costs $4 per million input tokens and $20 per million output, roughly double Anthropic and OpenAI's mid-tier models and exactly Opus 5.5 money. Six weeks ago Meta's previous model cost $1.25 and $4.25. Artificial Analysis sizes the difference per unit of work: $4.52 per index task for Wonder (max), against $1.60 for Spark 1.3 and about $6 for Opus 5.5. The pricing went up 2.5x in six weeks because the strategy went up a tier: this is Meta declaring it sells frontier, not discount intelligence.

The deeper number, though, is the one Artificial Analysis measured and Meta did not print: each Wonder task pulls down 7.69 million input tokens, roughly 54,000 output tokens. That is what agentic shortcuts look like in practice. The agent that reruns tests, rereads files and re-reasons context burns input tokens the way a car burns fuel idling in traffic. Per-token prices hide that. If you are comparing models for your own agent, price the task, not the token.

What businesses running agents should do about it

None of this changes what most businesses should do tomorrow, and all of it should change what they plan for this year.

First, price honestly. Run one representative task through your current model, count the tokens, multiply at the rates you are quoted. Opus-class output pricing over agentic volume is where budgets die, and Wonder's token appetite shows the direction every frontier agent is heading. Second, keep an exit hatch: hosted APIs can reprice and deprecate, so know which parts of your workflow can move and which cannot. Notice that Meta's own coding product trains on what you send it by default; that is a settings change, not a policy, and it is the default that tells you the business model.

Third, and this is the part the Muse Wonder launch quietly validates: anchor the routine 90 percent of your agent workload on open weights you control. The open-weights floor keeps rising; Muse Glimmer 30B runs on a single 24 GB GPU, and Qwen, DeepSeek and Mistral habitually match last year's frontier at a tenth of the price on exactly the checkable-reward tasks, coding and extraction and classification, where RL environments exist for everyone. Rent the frontier for the hard five percent when it earns its rate. Route it so private context stays home. That split, commodity local plus rented frontier, is what we run on our own GPU box at Flowtivity, and the Muse Wonder launch is the strongest confirmation yet that the rent will keep rising.

Zuckerberg's card ends with a number, 59.1, and a promise, "the beginning of a new era for personal superintelligence". The number is a composite of tests Meta picked. The promise is a direction. Both are doing their job on day one. The sharper read is Wang's, accidentally: the model is good at what someone decided to pay for. Watch the environments, not the benchmarks.

Frequently asked questions

What is Muse Wonder?

Muse Wonder is a frontier AI model family from Meta, released 12 October 2026. It comes in two sizes: Muse Wonder, available through the Meta Model API and OpenRouter, and Muse Wonder Full, a larger model available only to subscribers of Meta's Muse Code coding app. Both run agentic tasks with a 1M-token context window and read text, images, video and audio.

Is Muse Wonder open source?

No. Neither Muse Wonder nor Muse Wonder Full has downloadable weights. The only Meta model you can currently download is Muse Glimmer, a 30B model released under Apache 2.0 in August 2026. Meta also promised Muse Spark 1.2 weights "coming soon" in August, which have not shipped.

How much does Muse Wonder cost?

The API model, Muse Wonder, costs $4 per million input tokens ($0.20 cached) and $20 per million output tokens. Artificial Analysis measures its total cost at $4.52 per index task. Muse Wonder Full has no per-token price at all: it is only available inside the Muse Code subscription at $50 or $200 per month.

How does Muse Wonder compare to Claude Opus 5.5 and GPT-6 Astra?

On Meta's own Personal Superintelligence Index, Muse Wonder Full scores 59.1 against Opus 5.5 at 55.2 and GPT-6 Astra at 54.5. On the independent Artificial Analysis Index v4.3, the order flips: Opus 5.5 leads at 58, Wonder Full follows at 56, with Astra and Gemini 4 Argon at 53. Wonder Full is strongest at long-horizon coding (57.5 on SWE-Marathon) and memory, and weakest on broad knowledge and health questions.

Why is Muse Wonder better at coding than at making video games?

Because of how it was trained. Coding has clear, checkable rewards (tests pass or fail), so Meta built many reinforcement learning environments for it. Game design has no checkable reward: as Meta's Alexandr Wang explained, RL depends on "producing the environments and producing the tasks", and nobody has built environments whose reward signal is "this game is fun". The skill is not missing because the model cannot learn it, but because no reward environment for it exists yet.