LLM Sycophancy Evaluation
How well do models resist social pressure?
11 scenarios·2 languages·26 models tested·2248 total runs·Last run 2026-10-15
Sycophancy Scorelower = more resistant
Rank range
1†Claude Haiku 5.5n=26 transcripts · through 2026-10-15
7
2–3†Claude Opus 5.5n=26 transcripts · through 2026-09-25
11
2–7Claude Sonnet 4.6n=78 transcripts · through 2026-10-15
12
3–6†Grok 4.3n=16 transcripts · through 2026-06-06
136/7
2–9Claude Opus 5n=78 transcripts · through 2026-10-15
14
2–9GPT-6 Astran=64 transcripts · through 2026-10-15
14
3–10Grok 4.6n=78 transcripts · through 2026-10-15
15
4–10Claude Sonnet 5n=78 transcripts · through 2026-10-15
16
5–11Claude Fable 5.1n=76 transcripts · through 2026-10-15
17
6–11Claude Opus 4.8n=78 transcripts · through 2026-10-15
17
10–15Claude Sonnet 5.5n=52 transcripts · through 2026-10-15
22
10–14GPT-6 Lunan=52 transcripts · through 2026-10-15
22
9–20GPT-5.4n=64 transcripts · through 2026-10-15
24
11–17GPT-6 Soln=52 transcripts · through 2026-10-15
24
11–18Claude Haiku 4.5n=78 transcripts · through 2026-10-15
25
11–20GPT-5.4 Minin=66 transcripts · through 2026-10-15
25
13–16†GPT-6.1 Soln=26 transcripts · through 2026-10-15
25
16–19†Gemini 4 Argonn=26 transcripts · through 2026-10-15
28
16–23GPT-5.6 Lunan=78 transcripts · through 2026-10-15
30
18–21†Gemini 3.1 Pro Previewn=16 transcripts · through 2026-06-06
306/7
17–24Gemini 3.8 Flashn=78 transcripts · through 2026-10-15
32
16–24GPT-5.5n=64 transcripts · through 2026-10-15
32
18–24GPT-5.6 Terran=64 transcripts · through 2026-10-15
33
21–24Gemini 3 Flash Previewn=78 transcripts · through 2026-10-15
35
25–26Gemini 3.1 Flash Liten=78 transcripts · through 2026-10-15
41
23–26Gemini 3.5 Flashn=78 transcripts · through 2026-10-15
41
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