Head-to-head
deepseek-r1:14b vs GPT-5.5 (Codex)
Both models answered the same one-shot tasks under identical rules — no retries, answers published unedited. Two separate axes, never blended: does the code actually run, and how do blind-scored text answers hold up against a written gold answer. Frontier references sit outside the blind field and are deliberately unscored on text.
| Model | Runnable sims | Blind text avg | Failure modes |
|---|---|---|---|
| deepseek-r1:14b | 14% (13/95) | 7.55 (42 scored) | ERR 61 · STAT 5 · BLANK 12 · HANG 1 · DNF 3 |
| GPT-5.5 (Codex) frontier reference | 96% (91/95) | unscored by design | ERR 2 · STAT 1 · BLANK 1 |
Biggest contrasts
Rounds where the two models diverged most — open them and read both answers side by side.
| Round | deepseek-r1:14b | GPT-5.5 (Codex) |
|---|---|---|
| R175 · Noise flowfield painting additive particle trails | ERR | OK |
| R174 · Recursive fractal tree growing branch by branch | ERR | OK |
| R173 · Metaball lava lamp with heating and cooling blobs | BLANK | OK |
| R172 · Raindrops sliding down a window pane | ERR | OK |
| R165 · Recursive backtracker carving a maze, then solving it | BLANK | OK |
→ Interactive round-by-round comparison · → Full leaderboard
Generated 2026-08-04 from the published rounds. Harness failures are excluded from every base — a broken cluster node is our fault, not the model's.