Creative writing LLMs · dated snapshot

Which LLM writes long fiction best? What our dated 20-round test actually showed

No single model won across genres in our data. Foreverse, an Android and iPhone fiction reader, ran Fiction Bench on 2026-07-16: DeepSeek V4 Flash took first on Chinese xuanhuan fantasy but dropped to 6-7 on court romance, where DeepSeek V4 Pro led. Duels ran through 2026-08-13 on two Chinese novels only, with no English chains yet. Treat this as a four-month-old snapshot, because hosted models change.

Available for iPhone and iPad in selected regions outside mainland China. Not available on the mainland China App Store. Download help

Did one LLM beat the field across both fiction genres?

No, and the split between our two test books tells you more than an overall winner would. DeepSeek V4 Flash placed first in both blind rankings on traditional xuanhuan action fantasy, then sank to 6–7 out of 9 on court-intrigue romance. DeepSeek V4 Pro flipped that exact result: 2–4 on xuanhuan, 1–2 on court romance. A plain, fast, shouting style suits one book and fails the other.

Here is how the test worked. Nine models each wrote 20 consecutive continuation rounds on two Chinese novels—Ta Tian Jing (8.9 million characters) and Empresses in the Palace (210 chapters)—with every new round appended to the context window before generating the next. Temperature was 0.7. Two blind-ranking AI judge agents, each given a differently shuffled anonymous mapping, ordered the outputs while structural metrics cross-checked them. We build Foreverse and ran this pipeline ourselves, including a prompt-wording bug of our own on the Gemini row that we left visible on the board.

How did the nine models rank on each novel?

Numbers below are blind-review rank positions rather than point scores; ranges like 2–4 mean our two shuffled judge packets didn’t separate those models. DeepSeek ran on its official API, Gemini and Claude went through an aggregator gateway, and the rest ran through a pass-through gateway.

ModelXuanhuan fantasyCourt romance20-round failure seen
DeepSeek V4 Flash16-7None in 20 rounds
DeepSeek V4 Pro2-41-2None
GPT-5.6 Terra2-41-3Ending rewind on xuanhuan, rounds 18 to 20
GLM-5.22-42-3None; half-width quote marks throughout
Claude Opus 4.85-64-5None; punctuation drift late in the chain
Qwen3.7-Max5-68None; constant sensory and scent descriptions
Kimi K2.675-7Mid-chain collapse, 97.6% repetition at round 9
Grok 4.589Mid-chain collapse, looping passages
Gemini 3.1 Pro9*4-6Restart loop caused by our prompt wording

*Gemini’s 9th-place finish on xuanhuan came from a bug in how we phrased the prompt: under our original wording it rewound to the opening scene 20 times in 20 rounds, and after our fix it rewound 0 times. We kept the original rank and noted why. Price didn’t predict style fidelity either: in that July list-price snapshot, DeepSeek V4 Flash cost $0.14 per million input tokens—about three cents per 10,000 characters of new prose, or roughly one-fortieth the price of the most expensive model on the board.

What breaks over 20 chained rounds that a single prompt hides?

One-shot prompts never catch multi-chapter breakdown. Over 20 chained rounds we saw three distinct failure patterns: a restart loop that rewrites the opening every turn; a mid-chain collapse that repeats earlier passages (Kimi K2.6 hit 97.6% cross-round repetition at round 9); and an ending rewind, where GPT-5.6 Terra wrote 17 clean xuanhuan rounds before replaying chapter one across rounds 18–20, even while its court-romance chain ran without a single glitch. Mimicking an author’s prose tics and holding plot continuity across 20 turns turned out to be separate skills.

Text statistics and the blind judges diverged too. Qwen3.7-Max matched the source novels most closely on sentence length, dialogue ratio, and stock-phrase frequency, yet landed 5–6 on xuanhuan and 8th of 9 on romance because it kept inserting descriptions of scent into books that never use them.

What changed in the July and August follow-up duels?

As newer builds shipped, we ran head-to-head blind duels against their predecessors. Vote splits show xuanhuan first, then romance.

Challenger vs incumbentTestedXuanhuanRomanceWhat the duel showed
Kimi K3 vs K2.62026-07-1810:011:1Eliminated K2.6's whole-passage repetition collapse.
Qwen3.8-Max-Preview vs 3.72026-07-217:511:1Big romance improvement; slimmer edge on xuanhuan with plot-level looping.
Gemini 3.6 Flash vs 3.52026-07-239:31:10Opposite swings: redistributed failure modes across genres.
Grok 4.6 vs 4.52026-08-1312:012:0Cured both of 4.5's collapse patterns.
DeepSeek V4 Pro GA 0813 vs 0716 preview2026-08-133:91:11The release build lost to its delisted preview.

That bottom row is the practical lesson for anyone picking an API model. DeepSeek swapped the weights behind the official model ID, so the preview that won court romance on our board can no longer be called. Checking DeepSeek’s pricing docs on 2026-11-15 showed the same pattern on Flash: the old deepseek-v4-flash name is retired and now routes to DeepSeek-V4.1-Flash-1017, listed at $0.15 per million input tokens off-peak on a cache miss. Our July Flash rank describes weights that no longer answer under that alias.

What does this benchmark leave out?

Sample size is narrow: one book and one starting point per genre, one 20-round chain per model. Wide rank gaps are believable, while neighboring ranks lean on the two judge packs agreeing, and both judges are AI agents that may share a base model and its blind spots. We ran zero English chains, so nothing here measures English fiction. Nor does it test coding, reasoning, poetry, scripts, or single-turn flashiness, and it assigns no absolute scores. Because hosted models are moving targets, every rank stays tied to its test date and gateway.

Cost figures convert published list prices mechanically, ignoring prompt-cache discounts, peak-hour pricing, and volume tiers, so they compare models without predicting a bill. To test models honestly on your own novel, copy the benchmark’s setup on a smaller scale: branch from the same sentence with two models and keep going for many rounds rather than one, since the three long-run failure modes show up late. Foreverse runs on your own API key across 60+ preset providers at provider list price, and the free tier holds 3 books.

Who should not use this page to pick a model?

  • You write in English or another non-Chinese language. We have not run English chains yet, so these ranks do not carry over.
  • You want to pick a model once and never re-test. Providers swap weights behind the same model name, as the 0813 duel proved.
  • You want a single score to rank models by. We publish blind rank ranges because one chain per book cannot support finer precision.
  • You write poetry, dialogue scripts, or marketing copy rather than multi-chapter web fiction continuations.

FAQ

What is the best LLM for creative writing?

No single model won both genres in our data. DeepSeek V4 Flash ranked 1st on Chinese xuanhuan fantasy, while DeepSeek V4 Pro ranked 1–2 on court romance where Flash fell to 6–7. Those results come from 20 chained rounds per model on two Chinese novels (tested 2026-07-16, duels to 2026-08-13) and do not cover English.

What is the cheapest LLM that writes good fiction?

In our July snapshot, DeepSeek V4 Flash led on value: $0.14 per million input tokens at list price—about three cents per 10,000 characters—while ranking 1st on xuanhuan and 6–7 on court romance. DeepSeek's pricing page now routes that name to a newer build (DeepSeek-V4.1-Flash-1017) at $0.15 off-peak, so test it on your own book before committing.

Is the Fiction Bench data still current?

It is a dated snapshot: main chains ran on 2026-07-16, the last duel ran on 2026-08-13, and no newer runs had been added as of 2026-11-15. Because hosted models change weights under the same ID, treat these ranks as a starting hypothesis and re-test on your own text.

Does this ranking apply to English creative writing?

We don't know. Both test novels are in Chinese, and we haven't run English chains yet. When the #1 xuanhuan model can drop to 6–7 on another genre in the same language, assuming results transfer to English isn't safe.

How can I compare models on my own novel?

Import your book into Foreverse, highlight the same sentence, and run continuations with two or three models on separate branches. Keep going for multiple rounds, since restart loops, repetition collapses, and ending rewinds showed up late in our tests. With your own API key, you pay the provider's list price with zero markup from Foreverse.

Sources