Muse Wonder Full vs Muse Wonder: Same Launch, Same Rate Card, Four Points Apart on the Only Board That Measured Both

Muse Wonder Full and Muse Wonder arrived in a single post on October 12, 2026, from Meta Superintelligence Labs, and almost every specification in that post applies to both. Both are described with a 1,048,576-token context window. Both are listed at $4.00 per million input tokens, $0.20 for cached input and $20.00 per million output tokens. Both require reasoning. Both take text, image, speech and video in and return text. And on the independent Artificial Analysis Intelligence Index, they are four points apart: Muse Wonder Full at 56, ranked fourth of 699 models measured, and Muse Wonder at 52, ranked ninth of 219.

The four points are the smallest real difference between two frontier checkpoints released on the same day by the same lab, and they are also the least decisive thing about the pairing — because only one of the two is reachable. Muse Wonder answers on the Meta Model API and inside Muse Code. Muse Wonder Full is described by Meta as Muse Code subscription only, at $50 a month for Power Usage or $200 a month for Maximum Usage. This is not a comparison between a strong model and a stronger one. It is a comparison between a stronger one you cannot call and a weaker one you can.

One announcement, two checkpoints, and one column that separates them

Read the launch material line by line and the two models track each other almost to the end. Where they diverge is a single column, and it is not capability.

• Access — muse-wonder-1.0 answers on the Meta Model API and inside Muse Code; muse-wonder-full-1.0 is Muse Code subscription only, with no metered path and no SDK that can be pointed at it.

• Price — the shared rate card of $4.00 in, $0.20 cached and $20.00 out describes a path that exists for Muse Wonder and does not exist for Muse Wonder Full, whose cost is a monthly plan of $50 or $200.

• Independent index — Muse Wonder 52, ninth of 219; Muse Wonder Full 56, fourth of 699. Four points, measured on the same public harness, which is the only place both appear on one scale.

• Output budget — the callable checkpoint's single-response ceiling is 943,718 tokens. The subscription checkpoint's ceiling is not published, because nobody outside Meta can count it.

• Vendor composite — Meta's own table puts Muse Wonder Full at 59.1 on its Personal Superintelligence Index and Muse Wonder at 56. On the Artificial Analysis Index row of the same table Meta lists Muse Wonder Full at 59.0, while Artificial Analysis itself publishes 56.

That last line is worth sitting with. The vendor's number for its own top checkpoint is three points above the independent measurement of the same checkpoint on the same named index — and the independent measurement of the other checkpoint, 52, is closer to the vendor's 59.1 for the first one than the vendor's two figures are to each other.

Four points is not the number that matters

On Meta's own long-horizon rows the two checkpoints come apart by far more than four points, and in both directions. All of the following are vendor-reported, produced in Meta's harness at maximum reasoning effort, with nothing independently reproduced.

• JobBench, long-horizon professional tasks — Muse Wonder Full 75.3, Muse Wonder 30.8. A 44-point gap between two models from the same release.

• Terminal-Bench 4.0-AA — Muse Wonder 85.6, Muse Wonder Full 85.5. The callable checkpoint edges the subscription one, by a tenth of a point.

• SWE-Marathon — Muse Wonder Full 79.0, Muse Wonder 78.8.

• MemoryBench — Muse Wonder Full 39.4, Muse Wonder 34.3.

• Harvey LAB-AA, Meta's legal agent suite — Muse Wonder 54.3, Muse Wonder Full 13.1. The subscription checkpoint finishes last, 41 points behind the one you can actually call.

So the pair is four points apart on the independent board, 44 points apart in Meta's favour on one row, and 41 points apart against it on the next. Those three numbers cannot all describe one capability difference, and the most parsimonious reading is that they do not: a 44-point swing between checkpoints that sit four points apart on a public harness is what a harness-sensitive evaluation of a rollout-shaped task looks like in an aggregate score.

The independent board's four points are, if anything, the more honest number and the less useful one — a single-run gap that size sits inside the noise band of the evaluation itself. Taken together with the composite table, the three rows point at the same conclusion in aggregate: these are two top-ten frontier checkpoints whose general capability measures close, whose long-horizon behaviour measures far apart, and one of which nobody outside Meta can run. What it cannot establish is which one wins on your workload, and it does not try to.

What the extra rung is supposed to buy

Meta names two capabilities for the series that are not benchmark rows at all: Active Memory, described as persistent state carried across sessions, and Active Focus, described as self-managed attention over long tasks. If either works as described, it is a different kind of difference from a score — it changes the shape of the application, because state that survives a session boundary is something you would otherwise build yourself, badly, in your own database.

Neither is demonstrated on a model card, and neither is reachable for testing, because the checkpoint that carries them is the one behind the subscription. That leaves a team with a specific and unglamorous problem: the plausible reason to pay $50 or $200 a month for Muse Wonder Full is a feature nobody outside Meta can evaluate, attached to a checkpoint that also reports a 13.1 on the legal agent suite against its sibling's 54.3.

Independent measurement fills part of the gap and only part. Artificial Analysis records Muse Wonder Full consuming roughly 140 million output tokens during its evaluation against an 88 million median, and Muse Wonder roughly 130 million against a median of 85 million. Both spend heavily to reach their scores, which is consistent with mandatory reasoning and tells you the cost of an escalation before you make it — but neither figure says anything about whether persistent memory works.

The decision is a throttle, not a purchase

Most comparisons of two checkpoints end in a procurement question: which one do we buy. This one cannot, because one side has no metered SKU. That reframes the whole thing into a routing question instead, and the routing question has an unusually clean shape here, because the cheaper option and the more expensive one are sold on completely different meters.

Muse Wonder is metered. Its cost scales linearly with tokens at $4.00 and $20.00 per million, with a 95% cache discount on the input side, which is why the independent cost per index task comes out near $4.52. Muse Wonder Full is not metered at all. Its cost is a flat monthly plan that does not move with volume, which means the marginal cost of one more request to the Full checkpoint is zero within a plan and infinite outside one. Those two economics want opposite traffic: metered pricing rewards keeping the expensive model off most requests, and flat pricing rewards pushing as much as possible through it once you have paid.

Which means the practical architecture is an escalation ladder rather than a model choice. The callable checkpoint handles the bulk of traffic at a rate you can forecast. The subscription checkpoint absorbs the subset where its long-horizon behaviour is alleged to differ — assuming you have a plan that admits the volume, and assuming the 44-point JobBench claim survives your own measurement. The measurement itself is the part to build first, because a four-point public gap and a 44-point vendor gap cannot both be right, and your workload is the only instrument that resolves it.

Two things about the rest of the field make that ladder cheaper to build than it sounds. The first is that all of these models bill the same way — reasoning tokens are output tokens at $20 per million, and the effort ladder runs from minimal to max and defaults to medium, so the setting you choose is a bigger line item than the model identifier you swap. The second is that the comparison set this launch invites is already routable: Claude Opus 5.5 at $4.00 in and $20.00 out and Meta's own Muse Spark 1.2 at $1.25 and $4.25 are both on OrcaRouter at their vendors' list prices, alongside more than 200 other models behind one OpenAI-compatible API. List prices pass through at 0% markup, so the arithmetic of the ladder does not change when the model does.

Neither Muse Wonder nor Muse Wonder Full is one of our routes. There is no meta/muse-wonder or meta/muse-wonder-full entry in the catalogue, Muse Wonder Full is not available outside Meta's subscription surfaces at any price, and we are not going to describe either as though it were. What a router contributes to this particular decision is the failover and the fallback: the escalation rung that is supposed to run rarely is exactly the rung where a provider hiccup is least tolerable and least visible, and automatic failover plus a routing DSL that can express the escalation as configuration rather than as application code is what keeps a rarely-used path from becoming an outage.

The version to remember

If the only thing you take from this launch is that Muse Wonder Full scores higher than Muse Wonder, you have the least actionable part of it. The two differ by four points on the only independent board that measured both, by a vendor-reported 44 points on one suite and a vendor-reported 41 points against it on another, and by a hard binary on access: one is an API you can call today at $4.00 and $20.00 per million, the other is a monthly plan with no per-token rate and no way in from an SDK.

The right move is to treat the four points as evidence that these are peers in capability and the access difference as evidence that they are not substitutes. Build against the callable checkpoint, build the measurement harness that would tell you your own gap, and treat the subscription as an escalation you switch on when the measurement justifies it — not as an upgrade you buy because the composite score is higher. And when a vendor publishes a 44-point spread inside a single launch, the useful question is not which checkpoint to standardise on. It is which number survives contact with your own evaluation set, and that question only has an answer if you have already built it.