Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7

The 'pelican riding a bicycle' test is Willison's personal aesthetic preference, not a standardized benchmark, and the comparison pits a quantized local model against a cloud API model at very different cost and latency profiles. The result may say more about Claude Opus 4.7's SVG or image-rendering quirks than about general capability.

Anthropic is clearly under competitive pressure from multiple directions right now. TechCrunch reported on April 17 that Anthropic just launched Claude Design specifically to help non-designers generate visuals quickly, which suggests the company is aware its image-adjacent capabilities need a dedicated product push. Meanwhile, OpenAI's upgraded Codex (covered here from The Verge and TechCrunch on April 16) is attacking Anthropic on the coding side simultaneously. A hobbyist benchmark showing a smaller Alibaba model outperforming Claude on a creative visual task fits that broader pattern of Anthropic being squeezed, but one informal test from one developer is thin evidence on its own.

If Willison or other independent evaluators run the same Qwen3 model against Claude Design outputs on a wider set of visual tasks and the gap holds, that would be meaningful. If the advantage disappears outside this single prompt, it's a quantization artifact or a prompt-sensitivity issue, not a capability story.