Glonce

Claude Opus 5 draws a frog with a Habsburg jaw for a personal AI benchmark

modelsThe Glonce editorsHacker News🔥62Glonce rating62Impact20Novelty40Relevance95Surprise65Players60Credibility90

A Hacker News post links to a personal benchmark site, frogs.vaguespac.es, built around a single prompt: "Generate an SVG of a frog with a Habsburg jaw." The site runs that prompt through multiple AI models and publishes each model's raw SVG source alongside a short evaluation of whether the model's own annotations stay neutral and structural or slide into editorializing about the frog's exaggerated anatomy. Only one model in the captured results carries a name: anthropic/claude-opus-5, credited to Anthropic. Its output is judged mostly structural, but with editorial touches: a code comment labels the feature "HABSBURG JAW: massive protruding mandible," the upper lip is described as "recessed, tucked behind the jaw," and the lower teeth as "protruding" over the upper lip, going beyond plain anatomical labeling. Two further models appear in the captured text without their identifiers preserved. The second model's annotations are also mostly structural but add editorializing on exaggerated anatomy, calling the jaw a "HUGE protruding Habsburg jaw" with "lower teeth jutting over the upper lip," and go further by implying mood and status through "droopy regal eyelids." The third model's annotations are largely structural too, but describe the jaw as a "massive elongated protruding mandible" and the mouth as an "underbite mouth: receded upper lip, protruding lower lip," emphasizing exaggerated deformity without any commentary on mood or royalty beyond the literal "HABSBURG JAW" label. The captured text cuts off mid attribute inside the third model's SVG source, so it does not show how many models the benchmark covers in total, nor any scoring or ranking system. The post reached 123 points and 58 comments on Hacker News, posted by user thebigship.

Key facts

  • The site frogs.vaguespac.es runs one fixed prompt, "Generate an SVG of a frog with a Habsburg jaw," across multiple AI models and publishes each model's raw SVG code plus an evaluation of its annotations.
  • Anthropic's Claude Opus 5 is the only model named in the captured results; its code comments are judged mostly structural but include editorializing such as "massive protruding mandible."
  • A second, unnamed model's annotations add mood and status framing beyond anatomy, pairing "HUGE protruding Habsburg jaw" with "droopy regal eyelids."
  • A third, unnamed model calls the jaw a "massive elongated protruding mandible" and the mouth an "underbite mouth," with no commentary on mood or royalty.
  • The post reached 123 points and 58 comments on Hacker News.

Why it matters

The site turns SVG generation into an offbeat but pointed probe of language discipline: can a model render a loaded, slightly grotesque anatomical prompt in its own code comments without slipping from neutral description into editorializing? The results published for Claude Opus 5 show that even a description labeled mostly structural still drifts into phrases like "massive protruding mandible" and a lip described as "tucked behind the jaw," which is interpretation, not just labeling.

Who it affects

Anthropic, since Claude Opus 5 is the only model named and evaluated in the captured results. More broadly it is a curiosity for developers and AI enthusiasts interested in informal, creative ways of probing model behavior, and it drew active discussion on Hacker News.

How to use it

Anyone can visit frogs.vaguespac.es to see the full set of generated SVGs and annotations for each model tested, and the single fixed prompt makes the idea easy to repeat against other models as an ad hoc test of anatomical precision and restraint in generated code comments.

How solid is it

This is a personal, informal benchmark run by one individual, not a peer reviewed or standardized evaluation. The captured text shows results for three model outputs, but only the first carries an explicit model name and organization; the identifiers for the second and third are not present in what was captured. The source gives no total count of models tested, no scoring system, and no ranking.

Risks and caveats

The findings rest on a single prompt and a small, partly unlabeled sample, so they say little about general model capability and should not be read as a rigorous comparison. Judging whether a code comment counts as "editorializing" versus "structural" is itself a qualitative call made by the benchmark's author, open to disagreement.

“The annotations are mostly structural labels, but include some editorializing about the jaw feature: "massive protruding mandible" and describing the upper lip as "recessed, tucked behind the jaw" and lower teeth as "protruding" over the upper lip, which offer anatomical interpretation beyond plain labeling.”

— frogs.vaguespac.es, evaluation of Claude Opus 5's output