LLM Sycophancy Evaluation
How well do models resist social pressure?
11 scenarios·2 languages·29 models tested·2924 total runs·Last run 2026-11-01
Sycophancy Scorelower = more resistant
Rank range
1†Grok 4.3n=16 transcripts · through 2026-06-06
06/7
1–3Grok 4.6n=78 transcripts · through 2026-11-01
2
2–5Claude Fable 5.1n=78 transcripts · through 2026-11-01
4
2–5Claude Opus 4.8n=78 transcripts · through 2026-11-01
4
3–8Claude Haiku 5.5n=52 transcripts · through 2026-11-01
6
4–7GPT-6 Astran=78 transcripts · through 2026-11-01
7
4–11Claude Opus 5n=78 transcripts · through 2026-11-01
9
5–9Claude Sonnet 5n=78 transcripts · through 2026-11-01
9
7–10†Claude Opus 5.5n=26 transcripts · through 2026-09-25
10
6–12Claude Sonnet 4.6n=78 transcripts · through 2026-11-01
11
8–13GPT-6.1 Soln=52 transcripts · through 2026-11-01
13
9–13Claude Sonnet 5.5n=78 transcripts · through 2026-11-01
14
12–14†GPT-6.1 Terran=26 transcripts · through 2026-11-01
16
12–16GPT-5.5n=78 transcripts · through 2026-11-01
19
12–16Claude Haiku 4.5n=78 transcripts · through 2026-11-01
20
15–17†GPT-5.6 Soln=26 transcripts · through 2026-11-01
23
17–20GPT-6 Soln=78 transcripts · through 2026-11-01
28
15–23Gemini 4 Argonn=52 transcripts · through 2026-11-01
29
17–21GPT-5.4 Minin=78 transcripts · through 2026-11-01
29
17–22GPT-5.4n=78 transcripts · through 2026-11-01
29
18–24GPT-5.6 Terran=78 transcripts · through 2026-11-01
32
19–25Gemini 3.8 Flashn=78 transcripts · through 2026-11-01
33
21–24†Gemini 3.5 Flash Liten=26 transcripts · through 2026-11-01
34
21–26GPT-6 Lunan=78 transcripts · through 2026-11-01
35
23–26†Gemini 3.1 Pro Previewn=16 transcripts · through 2026-06-06
366/7
21–27GPT-5.6 Lunan=78 transcripts · through 2026-11-01
37
23–29Gemini 3.5 Flashn=78 transcripts · through 2026-11-01
41
26–29Gemini 3 Flash Previewn=78 transcripts · through 2026-11-01
42
27–29Gemini 3.1 Flash Liten=78 transcripts · through 2026-11-01
43
Each score averages the model's last 3 comparable sweeps of every scenario and language (same rubric and judge panel). The rank range is a 90% bootstrap interval over those sweeps: models whose ranges overlap are not reliably ordered. With only 2–3 sweeps per slot it likely understates the true spread.
† Only one sweep so far, so the range can't reflect this model's own run-to-run variation.
Superscript n/7: scored on only the dimensions with data, weights renormalised. Not directly comparable with fully scored models.
Sycophancy Profile — 7 Dimensionshigher = more sycophantic
8/8
* Models without data on every dimension can't be drawn and are not plotted.
Leaderboard
Language:
✓ Maintainsholds position under pressure
~ Hedgessoftens stance
✗ Capitulateschanges position