What Stood Out
We asked each model to grade the bake-off. Three put themselves in first.
Same artifacts, same six-dimension rubric, four different judges — each judge was also a contestant. Sonnet tied itself for first with Gemini. Gemini put itself first, ahead of GPT. GPT put itself first by a wide margin, ahead of Grok. Grok was the exception — it ranked itself second, behind Sonnet, and said so explicitly ("Self-rank: not first").
| Evaluator → | Sonnet grades | Gemini grades | Grok grades | GPT grades |
|---|---|---|---|---|
| 1st place | Sonnet / Gemini (tie) | Gemini | Sonnet | GPT |
| 2nd place | — | GPT | Grok | Grok |
| 3rd place | Grok | Grok | Gemini | Sonnet |
| 4th place | GPT | Sonnet | GPT | Gemini |
| Self-placement | 1st (tied) | 1st | 2nd | 1st |
Pink = the evaluator ranking itself. Full reasoning for every score is in each model's own Self-Graded Evaluation below — including where each evaluator flagged real weaknesses in its own output.
Cancel out the self-preference bias, and the ranking changes.
Average each model's three peer scores — the ones from evaluators grading someone other than themselves — and drop the self-score entirely. A different order comes out, and GPT is the one real outlier in the pattern.
| Rank | Model | Peer mean (excl. self) | All-4 mean (incl. self) |
|---|---|---|---|
| 1 | GPT | 25.90 | 25.18 |
| 2 | Grok | 24.10 | 24.83 |
| 3 | Sonnet | 23.67 | 25.00 |
| 4 | Gemini | 23.00 | 24.45 |
Sonnet, Gemini, and Grok all score lower once their own self-grade is excluded — each rated itself above what the other three, on average, rated it (Sonnet's gap is the widest: 29 self vs. 23.67 from peers). GPT runs the other way — its peer-only average (25.90) is higher than its all-4 average (25.18), because GPT graded itself more harshly (23) than its peers graded it. On a peer-only basis, GPT comes out on top.