AI fiction leaderboard
Which AI writes the best fiction?
Every model writes the same 20 genre stories, three chapters each. A panel of AI judges scores every chapter, and readers like you vote blind on short scenes.
Model rankings
Scroll sideways to see all scores. Select a model for its category breakdown.
| # | Spicy? | Samples |
|---|
No models ranked yet
Models appear here after they finish the full 20-genre benchmark.
How to read this table
AI judge score (0–100). A panel of AI judges scores every chapter and each finished story against a shared craft rubric plus genre-specific criteria. Hover over a score bar for its 95% interval across the 20 stories. Samples opens every story the model wrote, with the judges' notes.
Story averages the judges' scores for plot, emotional engagement, nuanced and believable characters, world building, originality, and genre craft. Intelligence measures how well the writer followed the prompt, its own chapter plan, and its character profiles, and kept the story coherent and well paced, with earned changes in tone and character. Prose quality averages dialogue, show-don't-tell, and freedom from purple prose, forced metaphor, and amateurish writing. All three are 0–100 across all judges and completed stories; problem criteria are inverted so higher is always better. Partial runs show their story coverage and exclude missing stories from averages. Click a model's name to see every category.
Human rating. Readers compare two anonymous versions of the same scene, written by different models, and pick the one they prefer. Votes are combined with a Bradley–Terry model, the method behind chess-style ratings. 1,000 is the average model, and a 100-point lead means readers prefer that model's writing about 64% of the time. ± is a 95% interval. Ratings with fewer than 20 votes are marked provisional.
Prices are OpenRouter's listed rates in US dollars per million tokens. Writing cost is what the writer model actually cost in its benchmark run, including planning, story and scene attempts, and retries. A full benchmark targets 20 stories and 60 comparison scenes. Judge costs are not included.
The boxes above the table show the top three models in each category. Best Bang for Your Buck weighs each model’s AI judge score against its price. The price blends the model’s input and output rates in the proportion of input and output tokens it used in its benchmark run. Every tenfold rise in price costs 10 points, so a model must score 10 points higher to justify costing ten times as much. Prices below $0.10 per million tokens count as $0.10.