Most AI benchmarks are numbers. A model scores 88.7 on some acronym, another scores 91.2 — and unless you work in the field, neither number tells you anything about what these models can actually do.
This is the first task in a series of human-friendly benchmarks: small, real tasks where you don’t need a leaderboard to understand the result — you can just look at it.
The task
Every model gets the exact same prompt, once, with no retries and no cherry-picking:
Generate an SVG of a walrus playing a cello. The walrus must have its two long tusks and visible whiskers. The cello must have a correctly shaped curved body with two f-holes, four strings, and a bridge, standing upright on its endpin. The walrus must be clearly drawing a bow across the strings with its flipper.
Why this is hard: the model writes SVG markup — shapes and coordinates — as text, without seeing what it draws. Getting a recognizable walrus, a structurally correct cello, and a plausible interaction between them is a surprisingly honest test of whether a model can reason about space it cannot see. (The idea of a fixed absurd drawing task owes a debt to Simon Willison’s famous pelican on a bicycle.)
The checklist, if you want to score along: two tusks · whiskers · curved cello body · two f-holes · four strings and a bridge · cello upright on its endpin · bow in flipper, across the strings.
The results
Pick a model — the drawing below is its actual, unedited output. New models are added as they come out.
Takeaways so far
A few patterns emerged after the first batch of frontier models:
- The scene itself is basically solved. Every model that returned a complete SVG drew something you would instantly caption “a walrus playing a cello”. Compare that to the stick-figure pelicans of 2024 — spatial drawing-by-text has come a very long way.
- Physical contact is the universal failure point. Nobody fumbled the walrus or embarrassed themselves on the cello, but every single bow is doing something impossible: hovering in mid-air, fusing with a flipper, slicing through tusks. Models know what things look like; they still don’t quite know how bodies touch.
- Details migrate to plausible-looking spots. Two models independently drew the cello’s scroll as a decorative curl on the walrus’s head. When a model can’t fit a part where it belongs, it doesn’t drop it — it relocates it somewhere that looks vaguely intentional.
- Polish and correctness are different axes. The most beautiful renders were not the most accurate ones, and the most naive-looking drawing ticked nearly every checklist item.
- Finishing at all was a benchmark of its own. Some models never made it onto this page: one kept truncating the SVG mid-file, others simply timed out — and one returned markup so creatively commented that no renderer accepts it. Before a model can draw a walrus, it has to reliably answer — that’s a result too.
- Frontier vs open weights is a real gap here. Every frontier model produced a recognizable scene; in the open-weights batch, only the smallest one managed a walrus, and the rest fell apart before or at the drawing stage.
Editor’s picks
Best so far — Gemini 3.1 Pro. The cello actually looks like a crafted instrument — shaded, proportioned, believable. The flippers are planted right where they belong, and the spotlight shadows sell the whole scene. This is the one drawing you could put on a poster — even if the musician himself looks suspiciously like a bear who borrowed a walrus costume.
Worst so far — Llama 4 Maverick. Brown abstract mountains for a cello, a gray balloon for a head, a red stick hovering where a bow should be. Grok 4.5, the previous holder of this title with its slapdash barrel-cello, is hereby pardoned — next to this, the barrel looks like fine craftsmanship. (Dishonorable mention: GPT-OSS 120B, whose drawing no one will ever see — its SVG doesn’t parse.)
This section will be updated as more models run. Kimi K3 and GLM-5 keep timing out on our runner — they’ll join the page when their endpoints start answering.
