Meta Says Muse Wonder Full Beats Opus 5.5 at Sustained Agent Work. Meta's Own Table Says Something Narrower.

Muse Wonder Full is the higher-scoring checkpoint Meta Superintelligence Labs announced on October 12, 2026, and Muse Wonder is the callable sibling that shipped in the same post. Both appear in Meta's own evaluation table beside Claude Opus 5.5, GPT-6 Astra and Claude Fable 5.1, and the claim that travelled furthest from that table — that Muse Wonder Full beats leading frontier models on several tests of sustained agent work, and that it outperforms every other model on DeepSWE including Claude Opus 5.5 — does not reproduce against the numbers Meta itself published. That gap is the story, and it is more useful than the claim.

What the claim says, and what the table says

The widely shared summary reports two figures in particular: a SWE-Marathon score of 57.5% for Muse Wonder Full against 46.3% for a Claude model, and a DeepSWE result that puts Muse Wonder Full ahead of everything else in the field. Meta's published rows read differently on both.

• SWE-Marathon — Meta lists Muse Wonder Full at 79.0, Muse Wonder at 78.8, GPT-6 Astra at 74.1, Claude Opus 5.5 at 74.2 and Claude Fable 5.1 at 67.4. There is no 57.5 and no 46.3 anywhere in the row, and the margin over Claude Opus 5.5 is 4.8 points, not 11.2.

• DeepSWE 1.1 — Meta lists GPT-6 Astra at 59.1 and Claude Opus 5.5 at 59.6, both above Muse Wonder Full at 57.1 and Muse Wonder at 52.0, which ties Claude Fable 5.1. On Meta's own table Muse Wonder Full is third on this row, not first.

• Terminal-Bench 4.0-AA — Muse Wonder Full 85.5, Muse Wonder 85.6, Claude Opus 5.5 81.5, GPT-6 Astra 81.4, Claude Fable 5.1 79.0. This is a genuine lead for the Meta pair, and the callable checkpoint edges the subscription one.

• JobBench, long-horizon professional tasks — Muse Wonder Full 75.3, GPT-6 Astra 34.2, Muse Wonder 30.8, Claude Opus 5.5 30.6, Claude Fable 5.1 27.6. A 44-point gap between two checkpoints from the same lab, which is either a real split in training targets or a harness artefact.

• Vals Finance Agent V2 — Muse Wonder Full 59.1, Muse Wonder 10.0, GPT-6 Astra 8.6, Claude Fable 5.1 6.4, Claude Opus 5.5 4.2. Another enormous internal spread, and the direction is not uniform across Meta's rows.

Every one of those numbers is vendor-reported, produced in Meta's own harness at maximum reasoning effort, and none of it has been independently reproduced. That is not a reason to dismiss the table. It is the reason the first question about any number that circulated from it should be which harness produced it.

The 57.5 and the 46.3 are not on the table

Two possibilities fit what is actually verifiable. Either the summary is quoting a different evaluation — a differently named suite, a different reasoning-effort setting, a pass rate rather than the index score, or a harness run outside Meta — or the figures are wrong. Nothing in Meta's published material establishes either, and this piece is not going to assert that Muse Wonder Full scores 57.5% on SWE-Marathon when the vendor's own row says 79.0.

The right response is procedural rather than rhetorical. A benchmark name is not a specification. SWE-Marathon run at maximum effort against a 200-step budget, a 500-step budget and a wall-clock cap will produce three different numbers for the same model, and the same is true of any long-horizon suite. When a score arrives without a harness, an effort setting and a step budget attached, it is a point on an unknown curve. Ask for the harness before you plan capacity around the number.

The corrected version of the claim is narrower and still interesting: on Meta's own long-horizon rows, Muse Wonder Full leads the Claude and GPT models in the table by margins that range from 4.8 points on SWE-Marathon to 44 points on JobBench, and loses to them on DeepSWE 1.1 and on Harvey LAB-AA. That is a defensible statement. The version that says it wins everywhere is not.

Which of the two you can actually reach

The claim has a second, harder problem, and it is not about scores at all. The checkpoint with the headline numbers is not sold per token.

• Where each runs — muse-wonder-1.0 answers on the Meta Model API and inside Muse Code, the company's terminal coding agent. muse-wonder-full-1.0 is described by Meta as Muse Code subscription only.

• What Full costs — Meta's published plans run $50 a month for Power Usage and $200 a month for Maximum Usage. There is no metered path, no per-token rate you can buy, and no way to point an SDK at the higher-scoring checkpoint.

• The rate card that does not apply to it — Meta publishes $4.00 per million input tokens, $0.20 for cached input and $20.00 per million output tokens for the series. Those rates describe a path that exists for Muse Wonder and does not exist for Muse Wonder Full.

• Context and output — 1,048,576 input tokens on both, with the callable checkpoint's single-response output budget capping at 943,718 tokens.

• Input surface — text, image, speech and video in; text out. Video and speech going in without a transcription step ahead of them is the capability difference most likely to change an architecture.

So a team reading the launch claim and then looking at its own procurement options is comparing two different things: an API at $4.00 and $20.00 that they can call today, and a checkpoint whose access is a monthly purchasing decision. Any evaluation that mixes them is not a model comparison.

What independent measurement covers

Artificial Analysis has measured the reachable half. Muse Wonder (max) sits at an Intelligence Index of 52, ranked ninth of 219 models, at $4.00 in and $20.00 out with a 95% cache discount, which works out to roughly $4.52 per index task — the cheaper per-task figure of the two frontier checkpoints measured. Its card records about 130 million output tokens consumed during evaluation against an 85 million median, the 1M context window, and text, image, speech and video input against text output.

Muse Wonder Full (max) reads 56 on the same index, ranked fourth of 699, with no price published because there is no per-token rate to normalise against. Meta lists the same checkpoint at 59.0 on the Artificial Analysis Index in its own table. A three-point disagreement on the same named index is a small thing in isolation and a clarifying one in aggregate: the vendor figure describes direction of travel, the 56 describes a measurement.

The independent board also tells you what the vendor table cannot. Muse Wonder Full scores 56 and Muse Wonder scores 52 — four points apart on a public harness, against the 44-point JobBench spread Meta reports from its own. One of those two instruments is measuring something the other is not.

What the escalation actually costs

Muse Wonder's pricing is not a value position. $4.00 in and $20.00 out is the same list price as Claude Opus 5.5, and several times what the current open-weight reasoning models charge. What separates the two at identical rates is how many tokens each spends finishing a task, and on that line Muse Wonder is the leaner of the pair: roughly 130 million output tokens against Claude Opus 5.5's approximately 260 million in the same evaluation framework, which is why the independent cost-per-task figures come out near $4.52 and $5.98 respectively.

The lever on that cost is the effort ladder, not a mode switch. Reasoning is mandatory, the ladder runs from minimal to max and defaults to medium, and reasoning tokens bill as output tokens at $20 per million. There is no non-thinking path to fall back on for high-volume traffic. Combined with a 943,718-token output ceiling, the effort setting is the single line in an integration that decides whether this model is affordable on your workload.

And the argument for reaching for the Full checkpoint at all is thinner than the launch claim suggests, because the four independent points between the two checkpoints on Artificial Analysis are inside the noise band of a single-run evaluation. What separates them on the vendor table are the long-horizon, stateful, tool-heavy rows — and those are exactly the rows nobody outside Meta has reproduced.

Where a router earns its place here

Muse Wonder is not on OrcaRouter. There is no meta/muse-wonder entry in our catalogue and we will not describe one as though there were; Meta's own API and Muse Code are where it runs, and Muse Wonder Full runs nowhere except behind a Muse Code subscription.

The routing problem this launch creates is a measurement problem. Once one model in a comparison is subscription-gated and the other is metered per token, the only way to compare your own workload across them is to run the same tasks through your own harness and count tokens rather than pass rates. That is the shape of work a single endpoint is built for. OrcaRouter carries more than 200 models behind one OpenAI-compatible API, and the models this launch invites you to measure against — Claude Opus 5.5 at $4.00 in and $20.00 out, Meta's own Muse Spark 1.2 at $1.25 and $4.25 — are already on it at their vendors' own list prices with nothing marked up on our side. Provider list prices pass through unchanged, so a vendor repricing reaches us the same day it lands.

There is a second use specific to this series. Meta has shipped at least three frontier-adjacent models in five months — Muse Spark 1.2 in August, Muse Spark 1.3 in September, the Muse Wonder series in October — and each arrives with a vendor scoreboard and a mandatory-reasoning cost model. Running a fixed evaluation set against whichever of them you can reach, and swapping a model identifier rather than a stack, is the cheapest way to avoid re-learning the same lesson three times.

What the launch claim does not settle

• Whether Muse Wonder Full ever gets an API path. The higher-scoring checkpoint is walled behind a monthly subscription, which makes access to it a purchasing decision rather than an engineering one. Meta has not said it will open it.

• Which harness produced 57.5 and 46.3. Until that is answered, the two most-quoted numbers from this launch cannot be placed on a curve alongside Meta's own rows, and should not be quoted as though they can.

• Whether the 44-point JobBench spread between two checkpoints from the same lab survives an independent run. A gap that size is either a genuine two-profile split or an artefact of the harness, and only a third party settles which.

• Whether Active Memory and Active Focus hold up in production. Meta describes persistent state across sessions and self-managed attention on long tasks. Neither is demonstrated on a model card, and both are settled by weeks of production traffic rather than by a launch post.

The series is real and shipped, and the callable endpoint is live. What the launch claim got wrong is the scoreboard, not the model. The defensible position today is the narrow one: on Meta's own long-horizon rows, the Muse Wonder pair leads the Claude and GPT models in the table on some suites and trails them on others, the checkpoint with the bigger numbers is the one you cannot call, and the two most-quoted figures from the announcement do not appear on the vendor's own table at all.