$ tail -f stew
a few weeks after I ran the Claude side of this test, OpenAI shipped GPT-5.6 — three models this time, not four: Sol, Luna, and Terra. Sol's the one that goes further than the rest, with an extra-high effort tier the other two don't get and Claude has no equivalent for.
same prompt as before, no changes: a hedgehog playing a violin, SVG only, no reference image. I kept the shape on purpose — a violin means a bow against strings, four limbs, a body covered in spines. hard for a model to fake, easy for a person to eyeball.
nine more hedgehogs, not ranked against Claude yet — just OpenAI against itself, effort tier by effort tier: the OpenAI gallery →
Terra reads like the workhorse tier, Luna the lightweight, Sol the one OpenAI is actually proud of. next: how all six models — three from each side — stack up against each other.