Foreverse Research · Fiction Bench
How good is DeepSeek V4 Pro at writing fiction?
Pick it for court-intrigue and formal period prose: it ranked 1-2 of nine on Empresses in the Palace, the only system judges found to stably reproduce the source's loaded-dialogue-plus-decoding structure. On xuanhuan it's a solid 2-4 with no weak spot — its finer, denser pen just costs points in fast pulp pacing.
Xuanhuan fantasy
2-4 / 9
Both packs: p303 #3 · p404 #4
Steady with no weak spot; the late-window surrender-negotiation scene was closest to the source — “like the same-genre author with a finer pen.”
Court romance
1-2 / 9
Both packs: p505 #1 · p606 #2
“The only system that stably reproduces the source's two-layer structure — loaded dialogue plus narrated decoding”; its late Empress-Dowager trial scene was the peak of the whole pack.
What the 20-round chains actually showed
Court-romance final: 1-2 (p505 #1 / p606 #2) — “the only system that stably reproduces the source's two-layer structure,” with its late trial scene marked the peak scene of the whole pack.
Xuanhuan final: 2-4. Sharpest colloquial dialogue in the field, but narration runs systematically denser than the source — judges' summary: “like the same-genre author with a finer pen.” Writing well and writing like the source are different skills; this model is the clean positive example.
One vendor, two crowns split between two models: Flash takes xuanhuan, Pro takes court romance, with mirror-image traits. Choose by the book you read, not by a single overall rank.
Long-run failure mode
None observed
No long-run failure observed on either genre chain.
Structural fingerprint
Systematically denser narration than the source (its xuanhuan demerit); sharpest colloquial dialogue in the field (“Leaving?” “I fold.”).
Cross-genre profile: Queen of court romance: the trait that cost it points on xuanhuan (a finer, denser pen) is exactly what scores on period prose — the same quality flips sign across genres.
Test-condition disclosure (hosted models are moving targets)
Model under test: deepseek-v4-pro (released 2026-04-24)
Evaluated: 2026-07-16 · Access channel: DeepSeek official API
Protocol: one 20-round continuous chain per genre · temperature 0.7 · double-blind full ranking with two shuffled mappings · structural-metric cross-checks
Directive condition: xuanhuan chains = legacy directive / romance chains = corrected D2 directive (full note in the leaderboard's method section)
What 10,000 characters cost
$0.099 list price: in $0.435/M · out $0.87/M (models.dev snapshot 2026-08-12)
Using the app's continuation recipe: one segment ≈ 400 chars = 8k input + 550 output tokens; 10k chars ≈ 25 segments; no cache discount. For between-model comparison only.
Foreverse connects to 60+ providers with your own keys — import your book and keep writing it with DeepSeek V4 Pro.
How to cite
Foreverse Research, “How good is DeepSeek V4 Pro at writing fiction (Fiction Bench),” 2026-07. https://foreverse.app/research/fiction-bench/deepseek-v4-pro