Support

Benchmarks · Writing & design

GameFeel Reverse Evaluation v1.5.2

Star1

GameFeel's reverse evaluation (RE), made much harder in v1.5: how well the model gives feedback on game proposals, especially on balance. Score: the board's RE (combined), about two thirds blind runs and one third assisted runs, where the model first sees 39 worked examples; out of 100. Blind and assisted scores are kept in each row's details.

Quality 84 of 100 · Private test set · Has measured 9 of today's 10 default lineup models

35 models · 52 results · updated 2026-11-23 · kodasdream.github.io ↗

Leaderboard

Gemini 4 Neon medium

Benchmark quality

How the score works
  • Overall
    84

    84 of 100

    How useful it is for telling today’s models apart

  • Headroom20%
    68

    Top models average 18 of 100

    Can top models still improve on it?

  • Test set20%
    100

    Private

    Could labs train on the questions?

  • Recency15%
    100

    3 days old

    How new is it?

  • Grading15%
    85

    Checked automatically

    How are answers scored?

  • New models15%
    90

    9 of 10 top models

    How fast does it test new releases?

  • Efforts10%
    51

    9 of 35 models at 2+ efforts

    Does it test each model at several reasoning efforts?

  • Breadth5%
    88

    35 models

    How many models has it measured?

How each part was worked out
Headroom 68 of 100 · 20% · +13.6
From the spread of published scores and the room left above the top quarter of models, one score per model.
Test set 100 of 100 · 20% · +20
Private test set. The reverse-evaluation items and answer keys are unpublished: over 8,000 tasks from the unpublished games (748 in a semi-private subset), plus a released-games part rating changes to commercial games such as Balatro, weighted 2 of 13. Source
Recency 100 of 100 · 15% · +15
Released 2026-11-19 (from the publisher's pages, not yet checked by hand).
Grading 85 of 100 · 15% · +12.8
Each rating is scored against a ground-truth answer key by a formula (rating error, ranking and separation); no model judges.
New models 90 of 100 · 15% · +13.5
Has measured 9 of today's 10 default lineup models. Too few new models to time how quickly it adds them.
Efforts 51 of 100 · 10% · +5.1
Measured 9 of the 35 models that come in more than one reasoning effort at two or more of them.
Breadth 88 of 100 · 5% · +4.4
35 models measured.

Not counted in the score: 1 star · used by 2 published pages. Both lift it in the directory’s recommended order. Test set, release date and grading were researched from the publisher’s pages (2026-11-19); tell us if one is wrong.

More from this benchmark

2 published measures

Breakdown

darker is better within each column · click a column to sort

All results

52 rows
#ModelEffortScore
1Gemini 4 Neonmedium25.4
2Gemini 4 Argonhigh25.4
3Gemini 4 Carbonhigh23.3
4Fable 5.2medium20.9
5Gemini 4 Neonxhigh20.0
6Gemini 4 Carbonmedium18.0
7Fable 5.2max18.0
8Fable 5.2high17.9
9Opus 5.5xhigh17.9
10GPT-6.2 Astrahigh14.6
11GPT-6.2 Astramax14.4
12Gemini 4 Hydrogenxhigh13.6
13Opus 5.5max12.5
14Gemini 4 Hydrogenhigh11.8
15Gemini 4 Carbonlow10.6
16Gemini 4 Hydrogenmedium10.5
17GPT-6.2 Solmedium8.7
18Sonnet 5.5max8.2
19Opus 5.5medium7.0
20Sonnet 5.5xhigh6.5
21Qwen 4 Maxxhigh6.2
22Haiku 5.5xhigh6.0
23Qwen 3.8 Maxxhigh5.7
24DeepSeek V4.1 Flash 1017max5.3
25Haiku 5.5high5.1
26Sonnet 5.5medium4.3
27Qwen 3.8 Flashxhigh4.0
28GPT-6.2 Solhigh3.8
29Qwen 3.7 Plusmedium3.7
30Nemotron 3.5 Super VLmax3.6