How the test was run

Fairness came from keeping everything identical except the model:

• Same conditions. Every prompt was pasted word for word into each model's own chat website, and I took the first answer — one attempt, no coaching. Each model runs at the mode shown next to its name in the table: each site's own default, except DeepSeek: Expert + DeepThink is switched on there because that setting serves the DeepSeek-V4-Pro model — the answers in its column are V4-Pro’s.
• Coding, graded automatically. Each program was run against a set of hidden checks — 124 across the eighteen tasks — that hammer the edge cases. A task only counts as passed if every one of its checks passes, so a single overlooked corner case fails the whole task.
• Reading, graded against a key. The long-document answers were checked against a prepared answer sheet: exact facts, exact contradictions, and — just as important — no invented ones.
• Drawings, judged two ways. The free-form drawing gets two grades — a mechanical element checklist, and my own ranking with comments (Task 24 shows both). The strict one was scored by a script that verifies ten precise requirements.
• Reasoning and language, graded against an exact key. The character-counting questions have one correct number each; the poem is checked programmatically against all four of its rules (line count, first letters, word count per line, matching end word); the misdirection puzzles each have one correct final answer, checked exactly — no partial credit for "close."
• The checks stay private so no model can "study to the test." The prompts and the answers, though, are all here for you to inspect.