Fiction Eval leaderboard · genre
Best AI model for fantasy writing
July 2026 — scores recomputed using only the fantasy scenes (84 blind matchups, three tries per model). small sample — treat as rough
Claude Sonnet 4.6 leads, and comes out ahead of #2 in 85% of re-runs. This fantasy-only sample is small, so treat the exact order as rough and check where the ranges overlap before reading much into close neighbors.
| # | Model | Fantasy score (likely range) |
|---|---|---|
| 1 | Claude Sonnet 4.6 | 1713 (likely 1664–1760) |
| 2 | Claude Opus 4.8 | 1682 (likely 1648–1707) |
| 3 | Grok 4.3 | 1682 (likely 1634–1732) |
| 4 | GLM 5.2 | 1652 (likely 1592–1696) |
| 5 | Gemini 3.1 Pro | 1436 (likely 1368–1505) |
| 6 | Mistral Large | 1406 (likely 1342–1470) |
| 7 | DeepSeek V3.2 | 1275 (likely 1209–1356) |
| 8 | Llama 4 Maverick | 1155 (likely 1102–1242) |
Claude Sonnet 4.6 on the magic-cost scene scene
The boy's breathing had gone wrong in a way Maren recognized. Shallow. Gapped. The fever had crossed into the chest. His mother stood at the door of the healing room with her hands pressed together like she was trying to keep herself from flying apart. A farmer's wife. Maren didn't know her name. They'd come from the eastern holdings, three hours by cart, and the boy had worsened on the road. Maren laid her palm on his sternum and felt the wet rattle of it. She had the spell. She'd carried it for six years, folded into her practice the way you fold a knife into a boot. She knew its cost too. The tax was proportional — a fever this deep, this settled, would take something substantial. Not a fact. Not a skill. Something lived-in. She stood and went to the window. The rule she'd made for herself: only trade your own. Never for a stranger's child what you wouldn't trade for your own blood. She had no children. …
Read every model's take side by side on the board.
More
The full 8-model board · the July 2026 report · all head-to-heads
How it works: every pair of models is judged blind on the same scene, with the passages'
order flipped so being shown first can't sway it, by GPT-5.4 — a family that isn't on the
board, so nobody scores their own side. Each score carries a likely range; overlapping ranges
are called a tie. Slop is scored separately by a fixed checklist, not an AI. Full board,
every prompt, and the FAQ: thebookfactoryai.com/board.
Get each new board by email on the model-drop list.
Writing a book of your own? Book Factory runs the same craft checks on full manuscripts.
Writing a book of your own? Book Factory runs the same craft checks on full manuscripts.