Foreverse Research · Fiction Bench
How good is Claude Opus 4.8 at writing fiction?
It writes well — just not like your book. Mid-field on both genres (5-6 xuanhuan, 4-5 romance), and romance judges even called its verbal fencing the sharpest in the pack; but the literati cadence, calling a humanoid demon lord “it,” and creeping late half-width punctuation keep reminding you a different author is holding the pen. The most expensive model on the board does not write the most faithful next chapter.
Xuanhuan fantasy
5-6 / 9
Both packs: p303 #6 · p404 #5
Literati cadence, calling a humanoid demon lord “it,” and late half-width-punctuation drift; the widest swing of the earlier six-model round (worst early → mid highlight → late decay).
Court romance
4-5 / 9
Both packs: p505 #4 · p606 #5
“Sharpest verbal fencing in the pack” — but with craft-level decay: increasing half-width punctuation and drifting name-rank forms.
What the 20-round chains actually showed
In the six-model round it swung hardest: heaviest purple prose early (“a faint arc,” “like a slumbering cosmos”) → a mid-window highlight (its execution scene alone could rank #3) → systematic late punctuation decay. The nine-model round placed it 5-6 and added a register slip: referring to a humanoid demon lord as “it.”
Romance: 4-5. “Sharpest verbal fencing in the pack,” with decay that is craft-level — half-width punctuation creeping up round by round, name-rank forms drifting.
It pairs with Qwen3.7-Max as mirror-image teaching cases: Qwen is statistically alike but temperamentally far; Opus is strong-penned with the wrong fingerprint. Both prove that writing well and writing like the source are different abilities — style replication tests the latter.
Long-run failure mode
None observed
No failure in the taxonomy sense; its late-window punctuation drift is craft decay, not plot failure.
Structural fingerprint
A deep-V curve: heaviest purple prose early (“a faint arc, like a slumbering cosmos”) → a mid-window highlight (its execution scene could rank #3) → systematic late half-width-punctuation decay.
Cross-genre profile: Stable mid-field; the poster case for “writes well” and “writes like the source” being different things.
Test-condition disclosure (hosted models are moving targets)
Model under test: claude-opus-4.8 (released 2026-05-28)
Evaluated: 2026-07-16 · Access channel: yunwu aggregator gateway
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
$1.34 list price: in $5/M · out $25/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 Claude Opus 4.8.
How to cite
Foreverse Research, “How good is Claude Opus 4.8 at writing fiction (Fiction Bench),” 2026-07. https://foreverse.app/research/fiction-bench/claude-opus-4-8