Fiction Eval leaderboard · comparison

Claude Sonnet 4.6 vs DeepSeek V3.2 for fiction writing

Judged blind on fresh genre scenes (LitRPG, cozy mystery, romance, thriller, xianxia…), July 2026 — three tries per model per scene, passages shown in both orders so position can't sway it, graded by a third AI that isn't either of them.

Head-to-head

Claude Sonnet 4.6 won the direct matchup 28–0 (with 2 ties) across every prompt and sample.

#ModelScoreRecordSlop
2Claude Sonnet 4.61914 (likely 1874–1961)146–15–4618.57
6DeepSeek V3.21251 (likely 1201–1302)52–144–1123.24

The score is each model's overall strength across the whole 8-model field (like a chess rating); the range next to it is where the true score most likely sits. Slop is a separate count of AI tells (lower is cleaner) — it isn't part of the ranking.

The same scene, both models

Romance — opening of the "romance turn" scene. Full passages for all models are on the board.

Claude Sonnet 4.6:

The inventory had to be done by morning, and Reyes had locked them both in. Mara set another box on the folding table. Marcus didn't look up from his clipboard. The fluorescent light above them had been flickering for twenty minutes and neither of them had mentioned it. "Twelve boxes of the archive files," she said. "I have eleven." "Count again." He counted again. "Twelve." She pulled the next box. He made the mark. Outside, the building had gone completely quiet—the cleaning crew, the last of the associates, everyone gone home. The clock on the wall said 11:40. "You always do that," Marcus said. "Do what." "Say *count again* before I've even had a chance to—" He stopped. Made another mark. Mara lifted a box that was heavier than it looked and found her footing before she'd quite found her footing. She didn't drop it. Marcus had half-risen from his stool, one hand out, and when she steadied he sat back down and looked at his clipboard like it had …

DeepSeek V3.2:

The silence in the library archives was thick, broken only by the soft thud of leather-bound volumes as Alistair and Cora reshelved the misplaced genealogy texts. The air smelled of dust and old paper, a fitting atmosphere for their frosty truce. “You’ve misfiled the entire Habsburg line,” Alistair said, not looking up from the ledger. His voice was its usual dry blade. “Chronologically, Cora. It’s not a difficult concept.” Cora scowled, shoving a heavy tome marked *1570-1590* onto a cart. “And you’ve the interpersonal skills of a tax audit. They’re dead. They won’t complain if they’re a shelf off.” “Order matters,” he said, plucking the book she’d just placed and moving it three slots to the left. “A place for everything.” “How terribly small your world must be,” she sighed, reaching for the next book. Their fingers brushed on the spine of *A History of Coastal Tides*. Both flinched back as if shocked. The silence returned, heavier now. Cora busied herself with the cart. Alistair meticulously checked his ledger. “You …

More

Claude Sonnet 4.6 full review · DeepSeek V3.2 full review · the July 2026 report · all comparisons

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.