Muse Wonder vs Claude Opus 5.5 on SVG Generation: What the First Hands-On Test Actually Shows

Muse Wonder went generally available on October 12, 2026 — the same day its higher-scoring sibling Muse Wonder Full shipped behind a subscription — and it has already accumulated more comparative pages than almost any model this blog has covered. What it had not accumulated, until this week, was a published first-hand test of a specific, checkable capability. On October 12 the X account @scaling01 posted a head-to-head of Muse Wonder against Claude Opus 5.5 on SVG generation, with two findings worth reading carefully: Muse Wonder Full is "good and even tries 3D generation," and its output style is "too gimmicky" and it "overindulges."

That is a real, specific, falsifiable claim about a real shipped model, and it comes from someone who reports having run both models themselves. It is also, at this writing, a single unreproduced run by one account whose own post opens with an advertising disclosure. Both of those things are true at once, and this piece is about holding them together rather than picking one.

The test's terms, stated as they were stated

The disclosure is the first line of the post rather than a footnote, which is worth noting because it is not typical of the format. The author states that Meta provided free preview access to Muse Wonder, that they paid for Claude Opus 5.5 out of pocket, and that the tests and conclusions are theirs. That is a materially weaker conflict-of-interest position than an undisclosed paid placement — one side of the comparison was comped, the other was not, and the author says so.

Three things about the setup are not given, and each of them constrains what the verdict can support.

• Which endpoint was tested — Muse Wonder and Muse Wonder Full exist as two separate model IDs with two different scoreboards, and the post names Muse Wonder in the comparison but attributes the 3D-generation observation to Muse Wonder Full

• Which effort setting was used — Muse Wonder's reasoning is mandatory and runs an effort ladder from minimal to max, defaulting to medium, and reasoning tokens bill as output, so the setting changes both cost and output character

• How many prompts were run — no prompt count, no prompt text, no seed, no rerun

The second and third of those matter more than they look. A model whose reasoning is always on and whose effort is user-set will produce visibly different SVG output at minimal than at max, and a verdict about style is exactly the kind of judgement that a single run of a single prompt can get wrong. This is not a criticism of the author — a short post on X is not a paper, and the disclosure plus the "tests and conclusions are my own" line is more transparency than most launch-day comparisons carry. It is a statement of what the piece can and cannot be used for.

What an SVG test actually measures

SVG is an unusually good target for a hands-on comparison and an unusually bad target for a casual one, and understanding why explains what the verdict does and does not tell you.

An SVG is text that renders deterministically. That is the good part: a reviewer can paste the model's raw output into a viewer, look at it, and see exactly what the model produced — no image-generation pipeline, no sampling seed on a diffusion decoder, no question about whether the criterion is the picture or the file. If a model emits malformed markup, the drawing fails visibly. If it emits 40 kilobytes of gradient definitions where 2 kilobytes of paths would do, that is measurable.

The bad part is that style verdicts like "gimmicky" and "overindulges" are the least separable claims in the whole space. Two independent judgements sit inside them:

• Craft — did the model produce valid, well-structured, renderable SVG, or did it produce something that needed repair before it drew anything?

• Restraint — did it use the minimum markup the request needed, or did it reach for gradients, filters, blurs, animation and embedded raster effects because they were available?

Craft is close to objective and can be scored by a script. Restraint is a taste judgement dressed as a technical one, and reasonable reviewers disagree about it constantly — a designer asking for a portfolio illustration wants the gradients, and an engineer asking for an icon sheet does not. So when a reviewer says a model overindulges, the useful question is not whether that is true but for which task it is true. The post does not separate the two claims, and it is genuinely hard to separate them in a single sitting.

The comparative claim, and the compliance claim

The post's framing puts Claude Opus 5.5 on the other side and the observation about 3D generation on the Muse Wonder Full side, which makes it a three-way comparison in a two-name title — the same structural wrinkle that runs through the whole Muse Wonder launch.

On that axis, the useful context is the same across all its siblings: Claude Opus 5.5 is not a brand-new model. It shipped on September 22, 2026, roughly three weeks before Muse Wonder, and its cost and capability profile is well characterised. Independently, on the Artificial Analysis Intelligence Index, Claude Opus 5.5 (max) reads 58 — the highest score in the 219-model class — while Muse Wonder (max) reads 52, ranked ninth of 219. Opus 5.5's list price is $4.00 per million input tokens and $20.00 per million output tokens; Muse Wonder's list price is identical, $4.00 and $20.00. The wit of the situation is that Opus 5.5 turns out to be the more expensive model in practice, not because of its rate card but because it is verbose: Artificial Analysis records it consuming about 260 million output tokens during evaluation against an 85 million median, against roughly 130 million for Muse Wonder, which works out to about $5.98 per index task against about $4.52. That is the kind of two-number comparison a token-metered evaluator can measure and a single SVG prompt cannot.

What the SVG test adds is a category the index does not cover. Structured-text generation sits somewhere between code and illustration, and no leaderboard scores it — not the Personal Superintelligence Index, not JobBench, not Terminal-Bench 4.0-AA, not SWE-Marathon. If a reader's actual job involves emitting clean vector files, the mainstream scoreboards are silent on their question, which is precisely why a scrappy head-to-head is worth linking at all.

What the second claim is really about

"Even tries 3D generation" is the more interesting half of the observation, and it is easy to misread as a capability claim.

In the SVG context, this most plausibly describes a model reaching past flat two-dimensional drawing into an implied or partial third dimension — isometric constructions, extruded shapes, perspective grids, stacked pseudo-layers, or generated markup that renders as a three-dimensional scene. It is not the same thing as a 3D model. Producing a mesh, a scanned object or a printable asset is a different output type entirely, and the phrase "3D generation" is used loosely across the space to mean anything from a 2.5D illustration to an actual geometry file. Without the prompts or the files, the reader cannot tell which one this was.

If it was the looser sense, the claim is consistent with the style complaint rather than in tension with it. Reaching for isometric projection or stacked layers is the same instinct as reaching for gradients: more machinery than the brief asked for. Both statements may be describing one behaviour — a model that treats available markup as permission — which is a more useful takeaway than either half on its own, and also the version a second reviewer is most likely to either confirm or break.

There is one asymmetry worth naming here, though. Where the two checkpoints differ in the launch data, the split runs along structure rather than polish. On Meta's own benchmark table, Muse Wonder Full beats the API model by a wide margin on JobBench (75.3 against 30.8) and Vals Finance Agent V2 (59.1 against 10.0) — while losing Harvey LAB-AA to it by 41 points, 13.1 against 54.3. Those numbers are vendor-reported, produced in Meta's harness, and unreproduced. But they describe a lab that trained two profiles rather than one, and a style or restraint difference between the checkpoints is exactly the kind of thing that a structural split would produce. Which checkpoint the reviewer was looking at when they wrote "gimmicky" is therefore not a pedantic question — the post puts the two observations on different models.

Where the models can actually be reached

For anyone who wants to run this comparison themselves rather than take the verdict on faith, the practical question is where each model lives, and the answers are not symmetric.

• Claude Opus 5.5 — generally available since September 22, 2026, $4.00 per million input tokens and $20.00 per million output, with a cache-read rate of $0.20, a 1-million-token context window, and thinking budgets that include a no-thinking configuration

• Muse Wonder — generally available since October 12, 2026, on the Meta Model API and inside Muse Code, at the same $4.00 in and $20.00 out, with a 1,048,576-token context and an output cap of 943,718 tokens

• Muse Wonder Full — announced on the same day but not sold per token; it is reachable only through a Muse Code subscription, and has no published per-token rate card, which is why it has no price row here at all

The catch is Muse Wonder's reasoning configuration. It cannot be switched off. The effort ladder runs minimal, low, medium, high, xhigh and max, defaulting to medium, and reasoning tokens bill as output. Any head-to-head run against a model that can be told not to think is therefore comparing a throttled configuration against an unthrottled one unless the reviewer states the setting, and this post does not. That is the single most consequential gap in the reproduction, more than the missing prompt count.

How to reproduce it in an afternoon

The reason this test is worth extending rather than dismissing is that it is cheap to extend. A useful version needs four things that the original does not supply: a fixed prompt set of maybe a dozen vector tasks spanning icon, diagram, illustration and chart; a statement of the effort setting; the raw SVG files kept alongside the renders; and a scoring pass that splits craft from restraint instead of blending them. Twelve prompts at two effort settings across two models is 48 runs — a few dollars of output tokens against these rate cards, and a genuinely more useful artifact than any launch-day impression.

This is also the case where a router pays for itself in a way a pure price comparison does not. Claude Opus 5.5 is served on OrcaRouter at Anthropic's own list price — $4.00 per million input tokens, $20.00 per million output, $0.20 per million for cache reads — with the zero-markup pass-through that applies across the catalogue, so a vendor price change lands on our side the same day it lands on theirs. Provider list prices pass through unchanged; the routing layer does not add a spread. A dozen-prompt sweep across two models is one key and one endpoint rather than two accounts, two SDK integrations and two billing relationships, and if you are sweeping something as unproven as a model released yesterday, automatic failover means a rate-limited or degraded provider does not take the whole run down with it.

To be explicit about what we do not have: there is no meta/muse-wonder or meta/muse-wonder-full entry in our catalogue. We do not route either checkpoint and we are not going to describe one as if we did. What we do carry from the same lab is the Muse Spark line — Muse Spark 1.2 sits at $1.25 per million input and $4.25 per million output, which is the same context window as the flagship at a fraction of the price, and it is the model to reach for if you want Meta's reasoning family through the same key as everything else. Adding Muse Wonder to a sweep is currently a second account; adding Muse Spark 1.2 is a line in the same config.

What to watch, and what would settle it

Three things would turn this from an interesting impression into evidence.

• A second reviewer with different taste reaching the same restraint verdict — style judgements that survive a change of reviewer are about the model; ones that do not are about the reviewer

• Muse Wonder Full becoming reachable per token — the model named in the 3D-generation observation is the one with no rate card, so nobody who wants to check that half of the claim can buy access to it without a monthly subscription to a coding agent

• Any independent scoreboard adding a vector or structured-graphics task — the gap this test falls into is not that nobody ran it, it is that no board scores it, which is why one person's afternoon is currently the whole record

The honest summary is that this is a real finding with a real limitation. A reviewer who disclosed paid-for preview access, paid for the other side themselves, and published a specific negative verdict has done more than a launch recap does — and a model that "tries" things it was not asked for is a claim about restraint, not about capability, and restraint is the easiest thing for the next reviewer to see differently. If you are choosing between Claude Opus 5.5 and Muse Wonder for vector output today, the independent index gap is six points and the token-volume gap is two to one in Muse Wonder's favour; the style verdict is one person's afternoon and should be treated as a hypothesis rather than a result.