FeelFish Evals · Human Leaderboard
This leaderboard is built on real readers: different models write the opening of the same novel task, and readers like or dislike the chapter after reading. Models are ranked by human like rate (likes ÷ (likes + dislikes)), with sample size, brief adherence and case coverage; models with fewer than 50 ratings are listed without a rank.
AI judge scores do not affect this ranking and are not shown here. They score prose, scene rendering, character, pacing and originality, and unlock on the model page after you rate a chapter. Like rate reflects reader preference, not a model's absolute writing ceiling.
At a glance
As of 2026-11-02, the top three by human like rate: xiaomi-mimo-2.6-pro (70.6%, n=102), claude-opus-5-5 (70.3%, n=101), gemini-3.8-flash (68.3%, n=101). See the table for full ranks, sample sizes and brief adherence.
Ranked by human like rate; models need at least 50 ratings to receive a rank.
This table is ranked only by human like rate. AI judge scores and sub-scores do not affect the order and are not published here; open a model page, read the first chapter and rate it to unlock them.
| Rank | Model | Human like rate | Likes / dislikes | Brief adherence | Latest eval |
|---|---|---|---|---|---|
| 1 | xiaomi-mimo-2.6-proChampion Xiaomi3 cases | 70.6% 102 ratings72 likes / 30 dislikes | 72/30 | 2/3 | |
| 2 | claude-opus-5-5Runner-up Anthropic3 cases | 70.3% 101 ratings71 likes / 30 dislikes | 71/30 | 3/3 | |
| 3 | gemini-3.8-flashThird Google3 cases | 68.3% 101 ratings69 likes / 32 dislikes | 69/32 | 1/3 | |
| 4 | Moonshot3 cases | 67.7% 99 ratings67 likes / 32 dislikes | 67/32 | 3/3 | |
| 5 | OpenAI3 cases | 64.3% 98 ratings63 likes / 35 dislikes | 63/35 | 1/3 | |
| 6 | ByteDance3 cases | 62.7% 102 ratings64 likes / 38 dislikes | 64/38 | 3/3 |
These models are already evaluated, but have not collected enough human ratings. Rates swing widely on small samples; read the sample size, case coverage and the chapter before drawing conclusions.
This is a reader-preference leaderboard, not a general capability score. These four notes matter more than the rank.
Like rate = likes ÷ (likes + dislikes), aggregated per model across all current published assessments. Each user can rate each assessment version once, which limits ballot stuffing.
Ranks go from high to low; ties are broken by sample size, then by the latest evaluation date.
Models below the threshold show their current rate and sample size but receive no rank; a handful of votes can swing the rate wildly.
Once the sample reaches the threshold, the model enters the main table automatically.
Only current (isLatest) published assessments are counted; old versions, drafts and unpublished records are excluded.
Models cover different cases and evaluation dates, so read cross-batch comparisons together with sample size and case coverage.
This page does not show AI judge scores and does not rank by them. They unlock on the model detail page after you rate a chapter, so machine scores are not mistaken for reader preference.
Like rate answers “do readers want to keep reading”; judge scores answer “how are prose, scene, character, pacing and originality”. They are not interchangeable.
By human like rate: likes ÷ (likes + dislikes), aggregated per model across current published assessments. Models need at least 50 ratings to receive a rank; ties are broken by sample size, then the latest evaluation date.
Models below 50 ratings are shown without a rank. With only a few votes the rate swings a lot, so a rank would mislead; they enter the main table automatically once the sample is large enough.
They answer different questions: like rate reflects whether real readers want to keep reading; AI judge scores rate prose, scene rendering, character, pacing and originality. The leaderboard uses only like rate; AI scores and sub-scores unlock on the model page after rating.
The page reads the latest votes and published assessments on every request, so blind PK votes and new assessments change the ranking. The page shows the data timestamp. Sample sizes differ by batch, so treat the board as a moving observation, not a permanent verdict.
Join a blind PK: read the anonymous openings and vote; identities and results unlock once enough participants join. To reproduce the setup, run the same brief and solution on the same novel-opening tasks in the FeelFish client.
Different blind PK batches attract different numbers of voters. We do not double-count votes or force different batches to the same sample size; read small samples with care.
