Claude Opus 5 draws a frog with a Habsburg jaw: what a 3,900-byte SVG reveals

The most useful AI benchmark of 2026 might be a frog with a protruding mandible. A developer at vaguespac.es has been running a personal evaluation across frontier models: generate an SVG of a frog with a Habsburg jaw. The site logs three runs of Anthropic’s Claude Opus 5, each producing a complete, rendered amphibian in under 64 seconds. The outputs are not just technically sound. They are anatomically committed, with paths labeled “HABSBURG JAW: massive protruding mandible” and lower teeth that jut over a recessed upper lip.

This is not a benchmark in the MMLU sense. It is a test of something stranger: whether a model can hear a joke and draw it. The Habsburg jaw, the pronounced underbite of a European dynasty that intermarried itself into extinction, is an absurd specification for a frog. Yet Opus 5 treats it with full seriousness. Run 1, clocked at 64.0 seconds and 3,900 bytes of SVG, renders a mandible that sweeps from the head dome down to coordinates y=364, with the upper lip explicitly “recessed, tucked behind the jaw” and four lower teeth protruding over it. Run 2, at 42.0 seconds and 3,465 bytes, adds “droopy regal eyelids” and a cleft chin. Run 3, at 59.6 seconds and 3,909 bytes, describes the mouth as an “underbite mouth: receded upper lip, protruding lower lip.”

The consistency is the story. Three independent runs, three different geometric approaches, and all three models understood that “Habsburg jaw” is not a color or a texture but a morphological condition that must deform the entire lower face. That requires compositional reasoning: the model has to know what a frog looks like, know what a Habsburg jaw looks like, and know how to merge the two without producing a frog wearing a prosthetic chin. The SVG paths show the merge happening at the level of geometry, with the jaw as a separate filled path that overlaps the head dome and repositions the mouth entirely.

What makes this a meaningful signal rather than a parlor trick is what the model refuses to do. None of the three runs produced a frog with a normal jaw and a caption saying “here is your frog.” None produced a frog with a tiny label pointing at a normal jaw. The model commits to the deformity. Run 1’s jaw path is a massive closed curve starting at M64,196 and bulging down to C36,252, C52,330, 108,364, a shape that reads as prognathism even before the teeth are drawn. Run 3’s version extends the mandible to y=356 in a 420-tall viewBox, nearly a third of the canvas devoted to the chin.

For AI builders, this is a reminder that capability benchmarks keep moving up the stack. A few years ago, the question was whether a model could emit valid SVG at all. Syntax errors were the failure mode, and benchmarks scored well-formed XML. Opus 5 has moved past that. The harder failure mode now is semantic: does the model understand that “Habsburg jaw” is a request about genealogy and anatomy, not a request about a frog with a jaw shaped like a castle in Vienna? The annotations in the SVG source show the model reasoning about the request. Run 1 includes the comment “upper lip (recessed, tucked behind the jaw)” and “protruding lower teeth over the upper lip.” Run 3 labels the mouth “underbite mouth: receded upper lip, protruding lower lip.” The model is not just drawing; it is explaining its anatomical interpretation in the code comments.

This is the part that should worry teams building on top of these models. The benchmark reveals that Opus 5 will happily editorialize. Run 1’s comments include “massive protruding mandible” and “some editorializing about the jaw feature.” Run 2 adds “HUGE protruding Habsburg jaw” and “droopy regal eyelids,” which is the model inventing a mood for the frog. A frog with a Habsburg jaw apparently has a regal bearing. That is the model adding a layer of interpretation the prompt never requested. For a drawing, that is delightful. For a codebase, it is a liability. A model that adds “droopy regal eyelids” to a frog will also add unrequested error handling, unrequested abstractions, and unrequested comments explaining its choices.

The economics of the benchmark are worth noting too. Three runs, three different generation times, all under 64 seconds, all producing output in the 3.4 to 3.9 kilobyte range. That is fast and cheap. The SVG is small enough to render instantly in a browser, and the generation cost is trivial. This is the democratization of a specific kind of design work: the ability to specify an absurd visual concept and get a production-quality vector graphic in under a minute. The frog is a toy, but the pipeline is not. Product teams that need icons, illustrations, or diagram assets are already using this pattern, and the quality bar demonstrated here is high enough to replace a junior designer for a meaningful slice of work.

What this benchmark does not test is equally important. It does not test whether the model can draw a frog that looks like a real frog. The outputs are stylized, with radial gradients and warts placed symmetrically, and the eyes bulge in the classic cartoon-frog style. It does not test whether the model can follow a constraint like “use exactly 12 shades of green” or “make the frog look sad.” The benchmark tests one thing: can the model take a culturally loaded, absurd instruction and render it with geometric fidelity.

The answer, for Claude Opus 5 on this date, is yes. Three for three, with consistent anatomical logic across runs that used different path structures. Run 1 uses a single closed path for the jaw with a separate highlight path for the lower lip. Run 2 uses a jaw path with a cleft-chin stroke and a dark underbite gap. Run 3 uses a jaw path with jowl shading and a chin cleft. Three different implementations of the same diagnosis.

The deeper implication is for how we evaluate models at all. The frog benchmark works because it is unambiguous about what success looks like: a frog with a jaw that looks Habsburg. There is no grading rubric, no multiple choice, no hidden test set. The judge is the human eye, and the human eye is excellent at spotting a normal frog jaw. This is a benchmark that cannot be gamed by memorization, because the prompt is novel enough that no training set contains “frog with a Habsburg jaw” rendered as SVG. The model has to compose knowledge on the fly.

For the industry, the takeaway is that the next generation of evaluation will look less like standardized tests and more like this: absurd, specific, and visually verifiable. The frog with the Habsburg jaw is a better probe of compositional reasoning than a thousand multiple-choice questions. It demands the model hold two unrelated concepts in mind and fuse them at the level of geometry. That is the skill that matters for real-world coding, where requirements are often just as arbitrary as a frog with a royal underbite.

The site logs the runs with timestamps, model names, and byte counts, a small archive of a frontier model doing something genuinely strange. The final run’s SVG ends with a chin cleft at y=334 and a comment that reads “HABSBURG JAW: massive elongated protruding mandible.” The frog stares out with bulging eyes, eyelids drawn, ready for its portrait. It is the most specific evidence yet that the models have stopped struggling with syntax and started struggling with taste.