Muse Wonder Is Live: Meta's New Frontier Model Ships With a Full-Tier Twin You Can't Buy Separately

Meta Superintelligence Labs released Muse Wonder on October 12, 2026 — and shipped it as two models in one launch. Muse Wonder is the checkpoint developers can call today through the Meta Model API and inside Muse Code; Muse Wonder Full is a higher-scoring configuration that Meta says is available only to Muse Code subscribers. Both report a 1-million-token context window and the same headline price of $4.00 per million input tokens and $20.00 per million output tokens, which puts the pair directly on top of Claude Opus 5.5 and makes the interesting question not what Muse Wonder scores but which of the two model IDs you can actually reach.

That split is the story of this launch. A frontier release normally gives you one endpoint and a number. This one gives you two IDs, two scoreboards, and an availability boundary drawn along a subscription line rather than an API tier.

What shipped, and where each model lives

Meta's own model page lists two IDs. muse-wonder-1.0 is the one served on the Meta Model API and through Muse Code, Meta's terminal coding agent. muse-wonder-full-1.0 is the one Meta describes as Muse Code subscription only — it is not a flag on the API model, it is a separate checkpoint reachable through a product Meta sells monthly. Meta's published plans for that product are $50 a month for Power Usage and $200 a month for Maximum Usage.

Both IDs are priced at $4.00 per million input tokens, $0.20 for cached input, and $20.00 per million output tokens. Meta also says Muse Wonder will "soon power experiences in Muse," which is the company's consumer surface — a rollout with no date attached yet.

Meta named two features in the release, and they are the ones worth tracking because they describe behaviour rather than a benchmark score. Active Memory is persistent state across a session; Active Focus is the model's own attention management on long tasks. Neither is measurable from a model card, and both are the kind of capability that only shows up after a few weeks in a real workload.

Meta's own scoreboard — and the two columns that matter

Meta published a benchmark table covering both checkpoints alongside Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5, all at their max reasoning settings. Every number below is vendor-reported, produced in Meta's harness, and none of it has been independently reproduced. Read it as Meta's claim, not as measurement:

• Personal Superintelligence Index — Muse Wonder Full 59.1, Muse Wonder 56, Claude Fable 5.1 52, GPT-6 Astra 53, Claude Opus 5.5 53

• JobBench (long-horizon professional tasks) — Muse Wonder Full 75.3, Muse Wonder 30.8, Claude Fable 5.1 27.6, GPT-6 Astra 34.2, Claude Opus 5.5 30.6

• Harvey LAB-AA (legal agent work) — Muse Wonder Full 13.1, Muse Wonder 54.3, Claude Fable 5.1 49.9, GPT-6 Astra 54.5, Claude Opus 5.5 55.2

• Vals Finance Agent V2 — Muse Wonder Full 59.1, Muse Wonder 10.0, Claude Fable 5.1 6.4, GPT-6 Astra 8.6, Claude Opus 5.5 4.2

• HealthBench Pro — Muse Wonder Full 62.3, Muse Wonder 56.6, Claude Fable 5.1 58.9, GPT-6 Astra 53.5, Claude Opus 5.5 58.6

• AutomationBench-AA — Muse Wonder Full 45.4, Muse Wonder 60.5, Claude Fable 5.1 62.1, GPT-6 Astra 65.3, Claude Opus 5.5 65.6

• GDP.pdf — Muse Wonder Full 33.2, Muse Wonder 39.1, Claude Fable 5.1 32.1, GPT-6 Astra 41.6, Claude Opus 5.5 41.2

• OSWorld V2.1 (computer use) — Muse Wonder Full 78.4, Muse Wonder 72.6, Claude Fable 5.1 73.1, GPT-6 Astra 66.1, Claude Opus 5.5 75.4

• MemoryBench — Muse Wonder Full 39.4, Muse Wonder 34.3, Claude Fable 5.1 34.3, GPT-6 Astra 34.3, Claude Opus 5.5 not reported

• Terminal-Bench 4.0-AA — Muse Wonder Full 85.5, Muse Wonder 85.6, Claude Fable 5.1 79.0, GPT-6 Astra 81.4, Claude Opus 5.5 81.5

• DeepSWE 1.1 — Muse Wonder Full 57.1, Muse Wonder 52.0, Claude Fable 5.1 52.0, GPT-6 Astra 59.1, Claude Opus 5.5 59.6

• SWE-Marathon — Muse Wonder Full 79.0, Muse Wonder 78.8, Claude Fable 5.1 67.4, GPT-6 Astra 74.1, Claude Opus 5.5 74.2

• Artificial Analysis Index — Muse Wonder Full 59.0, Muse Wonder 56, Claude Fable 5.1 54, GPT-6 Astra 53, Claude Opus 5.5 58

The Full column is not uniformly better, and that is the most informative thing in the table. Full wins JobBench by 44 points, Vals Finance Agent V2 by 49, and MemoryBench by 5 — long-horizon, stateful, tool-heavy work. Muse Wonder wins AutomationBench-AA by 15 points and GDP.pdf by 6, and it edges Full on Harvey LAB by 41. So the subscription model is not simply the API model with the dial turned up; Meta appears to have trained toward two different profiles, and the one you can call over an API is the stronger of the two on agentic business automation while the one behind the subscription is the stronger on sustained professional-task work.

What independent evaluation says so far

Artificial Analysis has Muse Wonder (max) at an Intelligence Index of 52, ranked ninth of 219 models, at $4.00 in and $20.00 out per million tokens with a 95% cache discount — working out to about $4.52 per index task. Its card reports 130 million output tokens consumed during evaluation against an 85 million median, 1M context, and text, image, speech, and video input against text output.

Muse Wonder Full (max) comes in at 56 on the same index, ranked fourth of 699 models, with no price published because it is not sold per token. The current leaderboard reads Claude Opus 5.5 (max) first at 58, then Claude Sonnet 5.5 (max) and Opus 5.5 (xhigh) and Muse Wonder Full (max) all at 56, with Opus 5.5 (high) at 54.

Note the gap between the two sources on the same model. Meta's table lists Muse Wonder Full at 59.0 on the Artificial Analysis Index; Artificial Analysis itself publishes 56. Meta's figure is Meta's; the 56 and the fourth-place rank are the evaluator's own. When a vendor and an independent lab disagree by three points on the same index, the independent number is the one to plan against.

The price is the same as Opus 5.5 — the bill is not

Meta's $4.00 in and $20.00 out is not a budget position. It is exactly what Claude Opus 5.5 costs, and it is several times what the open-weights reasoning models charge. At 130 million output tokens in an evaluation run, Muse Wonder sits in the middle of the field on verbosity: well below Opus 5.5's 260 million against a 85 million median, well above the leaner Gemini-class models.

Reasoning is mandatory on this model. The effort ladder is max, xhigh, high, medium, low, minimal, defaulting to medium, and reasoning tokens bill as output. There is no switch to turn reasoning off and no cheap non-thinking path — the only cost lever you have is the effort level and the size of the output budget, which caps at 943,718 tokens on a 1,048,576-token context. At $20 per million output, a long agentic loop that fills that budget is a real invoice; at medium effort on short calls it is a normal frontier-model bill.

The safety framing Meta used at launch is worth quoting because it is unusually direct: the company calls Muse Wonder its "most aligned model to date" and cites what it describes as its strongest safeguards yet against CBRN and cyber misuse. That is a vendor claim about its own model and should be read as one, but it does indicate where Meta expects the resistance to deployment to come from.

Where OrcaRouter fits — and where it does not, yet

Muse Wonder is not on OrcaRouter. There is no meta/muse-wonder entry in our catalogue and we will not describe one as if there were. What we do carry from the same lab is the Muse Spark line — Muse Spark 1.2 and Muse Spark 1.1 — so a team that wants Meta's reasoning family today has a route that does not involve a new account or a new SDK.

For the comparison shopping this launch invites, the useful fact is that both of Muse Wonder's pricing peers are already on our catalogue at their vendors' own list prices. Claude Opus 5.5 is on OrcaRouter at $4.00 in and $20.00 out with a $0.20 cache-read rate and the same 1M context, and DeepSeek V4 Pro is there at a fraction of that. If you are deciding whether Muse Wonder's $4/$20 is worth it for your workload, running the same tasks through two models that are already reachable costs you one API key instead of three.

That is the whole point of a single OpenAI-compatible endpoint across 200-plus models with zero markup: when a lab publishes a price like this one, you can measure it against the alternatives in an afternoon rather than opening a support ticket with three vendors. Provider list prices pass through unchanged, so a vendor price cut shows up on our side the same day it is announced. When Muse Wonder reaches a provider we route, it will appear at Meta's price, not at ours.

What to watch from here

Four things are unresolved about this launch, and each of them changes how you would plan around it:

• Whether muse-wonder-full-1.0 gets an API path. Right now the higher-scoring checkpoint is walled inside a $50-to-$200 monthly subscription, which makes it a product decision rather than an engineering one. Meta has not said it will open it.

• Whether the Active Memory feature survives contact with production. Persistent state across sessions is the claim that most changes how you architect an application, and it is also the claim a model card cannot demonstrate.

• Whether independent harnesses reproduce the JobBench and Vals Finance spreads. A 44-point gap between two checkpoints from the same lab on one benchmark is either a genuine capability split or a harness artefact, and only a third party can settle which.

• Whether the Muse consumer rollout arrives with the same model. Meta says Muse Wonder will "soon power experiences in Muse" without a date or a version, and the difference between a 1M-context agentic model inside a consumer app and a chat model wearing the same name is substantial.

The launch is real and the endpoints exist as of October 12. What is not settled is which of the two Muse Wonders is the one people will end up talking about — the API model at $4/$20 that anyone can call, or the Full checkpoint that currently requires a subscription to touch.