Fiction Eval leaderboard · report
The July 2026 Fiction Model Report
8 frontier models, 10 fresh genre scenes, three tries each, 819 blind matchups, judged by a different AI family, every score with a likely range.
Every model-launch cycle produces the same question in writing communities — "which model actually writes fiction best?" — and the same answers with no receipts. This report is the receipt: every ranking below comes from blind head-to-head matchups on scenes written fresh for this run (so they can't be in anyone's training data), graded by a model that never shares a family with either contestant, with a likely range on every score.
The board
| # | Model | Score | Record | Slop |
|---|---|---|---|---|
| 1 | Claude Opus 4.8 | 1927 (likely 1888–1976) | 150–14–43 | 14.88 |
| 2 | Claude Sonnet 4.6 | 1914 (likely 1874–1961) | 146–15–46 | 18.57 |
| 3 | Grok 4.3 | 1856 (likely 1815–1899) | 137–28–42 | 9.43 |
| 4 | GLM 5.2 | 1776 (likely 1732–1822) | 117–42–42 | 17.53 |
| 5 | Gemini 3.1 Pro | 1471 (likely 1423–1519) | 78–104–13 | 13.45 |
| 6 | DeepSeek V3.2 | 1251 (likely 1201–1302) | 52–144–11 | 23.24 |
| 7 | Mistral Large | 1118 (likely 1054–1167) | 35–162–10 | 17.79 |
| 8 | Llama 4 Maverick | 686 (likely 627–741) | 0–206–1 | 22.92 |
What stands out
- Claude Opus 4.8 and Claude Sonnet 4.6 are tied at the top. They're close enough that it's a coin flip — re-run the numbers and the two swap places about 33% of the time. Anyone telling you one of these two is definitively better at fiction is reading noise.
- Grok 4.3 is genuinely third — with three tries per scene it sits clearly below the top pair, not tied with them.
- Grok 4.3 writes the cleanest prose in the field (9.43 slop per 1,000 words) — but clean and good aren't the same thing, which is exactly why the board scores them separately.
- GLM 5.2 debuts at #4, above Gemini 3.1 Pro, in its first appearance on the board.
- Llama 4 Maverick lost every decisive matchup it played (0–206–1).
Method, briefly
Ten genre scenes (dungeon-core LitRPG, cozy-mystery cold open, enemies-to-lovers turn, air-traffic-control thriller, xianxia sect betrayal…), three passages per model per scene. Every pair of models is judged on the same scene twice — once in each order, so being shown first can't help — by a judge from a different AI family. All the wins and losses become one strength score (shown like a chess rating), and re-running the numbers many times gives the likely range next to each. A separate fixed checklist counts AI tells — the "it wasn't X, it was Y" crutch, essay words, stock phrases, filler — per 1,000 words.
Full excerpts, the blind-vote widget, the paste-your-prose AI-tell checker, and the FAQ are on the board. Genre-level boards: cozy mystery · fantasy · LitRPG · progression fantasy · romance · science fiction · thriller · cultivation / xianxia.
Writing a book of your own? Book Factory runs the same craft checks on full manuscripts.