Meta's Muse Wonder: Price Per Token vs Cost Per Task

Muse Wonder is Meta's first frontier model, launched on October 12, 2026 alongside Muse Wonder Full. Only Wonder is in Meta's API, at $4 per million input tokens and $20 output. Against Fable 5.1 and GPT-6 Astra the per-token price is lower, but per task the math changes: by Artificial Analysis's estimate, Wonder costs more than Astra. Here's the full math and the checklist for deciding.

What Meta launched: two models, not one

Almost everyone covered it as one model, but there are two. Meta Superintelligence Labs launched Muse Wonder and Muse Wonder Full:

• Wonder is in Meta's API and in Muse Code.
• Wonder Full isn't in the API: it only comes with the Muse Code subscriptions, $50 or $200 a month.
• Both take text, images, video and audio, answer only in text, and handle a million tokens of context.
• Neither has downloadable weights: you rent it, you don't own it.

It lands fourteen months after Meta started rebuilding its AI from scratch, and it's the first time Meta sits at the pricing table with the big players.

Active Memory and Active Focus

Meta is showing off two pieces. Active Memory: the model keeps what matters about you between conversations and pulls it back when the task calls for it. Active Focus: if something blocks it, it tries another path, and before an important action it warns you or asks.

The key idea is that Meta isn't only selling intelligence, it's selling persistence. A model that remembers you and doesn't give up at the first error changes how often you have to step in, and every time you step in is a cost the pricing table doesn't show.

How to try the API step by step

1. Go to dev.meta.ai and create your API access.
2. In your code, only change the model name to muse-wonder-1.0: the API is compatible with OpenAI and Anthropic tooling.
3. Send it the task with its tools. The model returns text; Wonder doesn't generate images or audio.
4. Where people get stuck the first time: you want Wonder Full and it's not there, because it isn't in the API. And if you ask for zero data retention, that option leaves out conversation state on the server, which is exactly what powers the memory.

You don't know it worked from the benchmark: run twenty of your real tasks and write down what each one cost. The per-token price is flat: it doesn't go up even if you fill the whole million of context.

The per-token math

Take an agent task with 200,000 input tokens and 20,000 output.

• Muse Wonder: $0.80 for input and $0.40 for output, or $1.20.
• Fable 5.1 or GPT-6 Astra: $2 for input and $1 for output, or $3.
• The difference: $1.80, 60% less. A thousand tasks a month: $1,200 versus $3,000.
• Opus 5.5: charges exactly what Wonder does, $4 and $20, so there's no per-token difference.

But that's per token, and what you pay for is a task.

Three cases (WAiBOT analysis using Meta's numbers)

Long coding tasks. A software agency with tasks that touch lots of files and get fixed over several rounds. On SWE-Marathon, Meta reports 53.1 for Wonder, 42.5 for Astra and 46.3 for Opus. If you use Opus, the per-token price is identical and the per-task estimate drops by almost a quarter. Those are Meta's numbers: the honest test is to move ten real tasks and count the correction rounds.

Legal drafts. On Harvey LAB-AA, Wonder passes 10% of tasks, Astra 8.6 and Opus 4.2. Sounds good until you read what it measures: a task only counts if it meets every criterion and invents nothing. In other words, nine out of ten deliverables still need a lawyer. The point isn't replacing anyone; it's getting the draft closer to signature.

An assistant with caching. A knowledge base of 500,000 tokens costs $2 to read without cache. If the read hits the cache, at 20 cents per million, it costs 10 cents: twenty times less. With that math, an assistant that rereads the whole file on every question stops being a luxury. What makes it hard to copy is that it depends on your data being organized before the model ever sees it.

Who it's NOT for

• Anyone who needs open weights to run the model at home: Meta didn't release them.
• Anyone who wants Wonder Full through the API: it only exists inside Muse Code.
• Anyone coding with confidential code in Muse Code: the default mode of those subscriptions lets Meta use your prompts and your code to improve its models.
• Anyone who needs fine-grained health answers: on HealthBench Pro, Wonder sits below Astra and Opus.

The objection: per token or per task?

"$4 versus $10 is 60% less, I'm switching today." You're right about part of it: against Fable 5.1 and Astra, per token, the savings are real. But you don't pay per token, you pay per task. In Artificial Analysis's estimate, Astra costs $3.26 per task and Wonder $4.52: almost 40% more. For Wonder to pay off against Astra, it would have to solve your task in fewer rounds.

The objection wins on terminal work: there, Astra and Opus score 59 against Wonder's 52. So don't migrate everything: pick one workload, measure a hundred tasks, and decide with your bill, not the table.

Where Meta's benchmarks come from

The claim going around is that Wonder and Wonder Full tie Fable 5.1, GPT-6 Astra and Opus 5.5. That tie comes from Meta's own spec sheet, and right there Meta explains how it picked the numbers: when there was no third-party result, it used the highest of its internal run, the self-reported figure, or the official leaderboard. The competitors' runs were also done by Meta, and may not be well tuned. And MemoryBench, where Wonder wins, is an internal test of Meta's own star feature.

Artificial Analysis already verified part of it: Wonder scores 52, Astra 53 and Opus 58. Very competitive, not superior. The sheet also confesses uncomfortable things: in ExploitGym it captured 92 flags, but only 21 with the intended vulnerability.

Where Muse Wonder breaks

• Meta admits that visual and spatial reasoning, especially in three dimensions, is still weak.
• It admits errors on multi-step tasks it hasn't seen much of, and that it didn't evaluate every language: if you work in Spanish, test with your own texts.
• It says performance changes with the harness, the prompt and the tools, so your result won't match the table.
• There's a limit of 3,000 requests and 4 million tokens per minute. That's enough, but with parallel agents it fills up fast.

How to apply it in your business?

The full math, using Artificial Analysis's estimate for a thousand tasks: Wonder $4,520, Opus $5,980 and Astra $3,260. Wonder saves $1,460 against Opus and spends $1,260 more than Astra, because per task it uses roughly three and a half times Astra's tokens. The checklist:

1. Measure cost per task, not per token.
2. With big contexts, the 20-cent-per-million cache makes everything cheaper.
3. If your code is confidential, use the API in standard mode and turn off Muse Code's contributor mode.
4. Put a person in front of any irreversible action.
5. Test a hundred of your own tasks before migrating.

And the warning that avoids the costliest mistake: Wonder Full isn't in the API; don't budget around a model you can't buy by the token.

To see how Meta's previous model stacked up, read the full breakdown of Muse Spark 1.3 against GPT and Opus.

Muse Wonder FAQ

How much does Muse Wonder cost?

$4 per million input tokens and $20 output in the API. Wonder Full isn't sold through the API, only in Muse Code ($50 and $200 monthly plans).

Is Muse Wonder cheaper than Opus 5.5?

Per token, no: they charge the same. Per task, Artificial Analysis's estimate puts it cheaper than Opus ($4,520 versus $5,980 per thousand tasks) and more expensive than Astra ($3,260).

Does it have open weights?

No. Neither model has downloadable weights.

Can I use Muse Wonder Full through the API?

No, only inside the Muse Code subscriptions.

Can I trust Meta's benchmarks?

They're self-reported. Artificial Analysis verified part of them and puts the model as very competitive, not superior.

Conclusion

Muse Wonder isn't a cheap model; it's a competitive model with Opus pricing, and that's a different conversation. The question that's left is whether Meta's active memory justifies switching providers, or whether it's just another pricing argument. Tell me which of your own tasks you'd measure first against Muse Wonder, and against which model.

Sources: Meta blog, model card, Muse Code, VentureBeat and TechCrunch.