Gallery, prompt v1
A caffeine molecule in 3D
One frozen prompt and every model that has run it, newest first.
Saturated since 2026-09-21
It is off the board. All five checks are measured, and both models passed all five. It was called on one run per model, so the same evening Qwen3.8 Max had three more runs and passed every check on all three. DeepSeek 4.1 Flash had two valid runs, both perfect, because two more were voided for reading the grader. On 2026-09-22 it had three more with the grader out of reach, and passed every check on all three. A textbook molecule is about what a textbook molecule should cost them. If a new model fails it, it comes back on the board.
The prompt, v1, frozen
Write a PDB file with 3D coordinates for a caffeine molecule, hydrogens included.
Every model, its first three runs.
One square a check, in this order:
- C8H10N4O2
- Fused purine core
- Methyls on N1, N3, N7
- Carbonyls at C2 and C6
- Flat rings, real bond lengths
passed, failed, undecided.
| DeepSeek 4.1 Flash | |||
| Qwen3.8 Max |
How it is checked
Caffeine, C8H10N4O2: a purine core, a six-membered ring fused to a five-membered ring across two shared carbons, methyls on N1, N3 and N7, carbonyl oxygens at C2 and C6, and the whole ring system flat.
What replaces it: Ask for the crystal conformer, where the methyl hydrogens have a defined rotation.
The rules behind all of thisThis task's prompt, checks and any checker, on GitHub
run 1 of 20 Sept, 3 rounds, $0.04The same five checks Qwen3.8 Max passed, right on the first render, for a fifteenth of the price.
run 3 of 21 Sept, 4 rounds, clock stop, $0.06Right on the first render and unchanged through four, until the clock ended it.
run 1 of 22 Sept, 4 rounds, $0.07Right in four rounds, most of them spent renaming atoms to the standard ligand names.
