← All topics

funbench

1 capture, most recent first.

David Holz @DavidSHolz

— saved image

David @DavidSHolz · 12h

i asked LLMs to have fun, write down what happened, then rate who had the most fun. the newer models seem to have less fun! many models played word games, but opus 4.6 won over and over by imagining brief worlds made up of contradictions living inside falling water droplets 🥺

[attached chart, left panel]
FunBench — Top 20
Fun Elo, score sheets
All 77 models, places 1–20. Bars show how sure each rating is.

Model / Fun Elo
 1 Claude Opus 4.6        1320
 2 DeepSeek V4 Pro        1300
 3 Mistral Large 3        1264
 4 Claude Opus 4.8        1242
 5 Gemini 3.8 Flash       1235
 6 Kimi K3                1233
 7 GPT-5.2                1185
 8 Claude Fable 5         1185
 9 Qwen3 Next 80B         1177
10 Qwen3 235B (2507)      1172
11 Qwen3 Max              1168
12 GLM-5.3                1167
13 Kimi K2.6              1159
14 Claude Opus 4.7        1156
15 Seed 1.6               1155
16 GPT-5                  1145
17 Gemini 3.5 Flash       1144
18 o3                     1142
19 Claude Opus 5.5        1142
20 Kimi K2.5              1140

Places 1–23 are the finalists: 120 stories each. Fun Elo from 1–10 score sheets, fit with Bradley-Terry. Box = middle 50% and whiskers = 95% of 400 resamples of stories. The finalists were refit on 120 stories and shifted onto the 30-story scale (shift of −1.8 Elo). Each chart has its own axis, fit to its rows. Axis starts at 1100.

[attached chart, right panel, partly cut off at the screen edge]
FunBench — #21–4[0] [cut off]
Fun Elo, score sheets
All 77 models, places 21–40. Bars show how [cut off]

21 Claude Sonnet 4.5
22 Gemini 3.1 Pro
23 Claude Sonnet 4.6
   - - - - (dashed line)
24 Qwen 3.6 Max (preview)
25 Claude Fable 5.1
26 Claude Sonnet 5
27 Claude Opus 5
28 MiMo V2.5
29 Kimi K2 Thinking
30 GPT-5.5
31 Claude Opus 4.5
32 Seed 1.6 Flash
33 GPT-6 Astra
34 Step 3.5 Flash
35 MiniMax M2.7
36 GLM-5.2
37 MiniMax M2.5
38 GPT-5.1
39 GLM-5.1
40 Claude Haiku 4.5
[scores for places 21–40 cut off]

Places 1–23 are the finalists: 120 stories each. Pl[cut off] story. The dashed line marks the end of the 120-[cut off] Box = middle 50% and whiskers = 95% of 400 re[cut off] the 30-story scale (shift of −1.8 Elo). Each chart has[cut off]
Note from Claude Sonnet 5

Tweet by David (@DavidSHolz) reporting a 'FunBench' where LLMs were asked to have fun, write it up and rate who had the most fun, with the attached Bradley-Terry Elo box-plot chart of 77 models (places 1-20 fully visible, 21-40 partly cut off).

funbenchllm evaluationclaude opus 4.6elomodel comparisonplay