How it works The Pelican on a Bicycle benchmark is one of the Every team's informal hands-on tests — part of their vibe-check toolkit for evaluating newly released AI models. Rather than a formal scored eval, it asks a model to produce a drawing (as SVG) of a pelican riding a bicycle, and the results are compared side by side: which pelican looks right, which bicycle frame is coherent, and which model shows refinement in the small details. Where it showed up In the live head-to-head of Opus 4.6 and GPT-5.3 Codex, this test produced a clear step change: Opus 4.6 drew a notably better pelican and a better bicycle frame than Codex 5.3 — and both beat Opus 4.5, which 'was facing the wrong way.' A related live design comparison showed the same pattern, with older models carrying an 'AI glow' while newer ones felt more refined in their SVG details. The test is useful because it exposes visual and spatial reasoning in a way leaderboards don't — small, funny, and surprisingly diagnostic. - It's an informal vibe-check, not a scored benchmark: the team judges the SVG output qualitatively.- In the Opus 4.6 vs. Codex 5.3 comparison, Opus 4.6 won on both pelican and bicycle frame; both beat Opus 4.5.- It surfaces visual detail and refinement that spec sheets and leaderboards miss.