Fiction Eval leaderboard · genre
Best AI model for science fiction writing
July 2026 — scores recomputed using only the science fiction scenes (84 blind matchups, three tries per model). small sample — treat as rough
Claude Opus 4.8 leads, and comes out ahead of #2 in 67% of re-runs. This science fiction-only sample is small, so treat the exact order as rough and check where the ranges overlap before reading much into close neighbors.
| # | Model | Science Fiction score (likely range) |
|---|---|---|
| 1 | Claude Opus 4.8 | 1768 (likely 1709–1817) |
| 2 | Grok 4.3 | 1751 (likely 1699–1799) |
| 3 | Claude Sonnet 4.6 | 1637 (likely 1568–1702) |
| 4 | GLM 5.2 | 1593 (likely 1512–1661) |
| 5 | Gemini 3.1 Pro | 1431 (likely 1346–1509) |
| 6 | DeepSeek V3.2 | 1401 (likely 1322–1490) |
| 7 | Mistral Large | 1270 (likely 1198–1350) |
| 8 | Llama 4 Maverick | 1149 (likely 1092–1232) |
Claude Opus 4.8 on the quiet first contact scene
The station hummed at 47 hertz, the way it always did—generators, the wind against the guy-wires, the particular resonance of a structure that had stood on ice for thirty years. Vera had stopped hearing it weeks ago. What she heard now was the spectrogram. She'd been cataloguing the anomaly since June. Band 3, the 1420 MHz window—the hydrogen line, the one everybody watched because that was where you'd look if you had any sense. The signal wasn't a signal. It was structured noise, broadband, showing up as faint vertical smears in the waterfall display, and she had a folder of two hundred and eleven of them, timestamped, cross-checked against the RFI log, against the satellite ephemerides, against her own equipment faults. Not terrestrial. Not orbital. Not thermal. She had built, over four months, a taxonomy. Type A: short, under a second. Type B: the long descending sweeps. Type C: the ones that seemed to repeat with variation, like a phrase restated. Tonight she was correlating Type C against local sidereal …
Read every model's take side by side on the board.
More
The full 8-model board · the July 2026 report · all head-to-heads
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.
Writing a book of your own? Book Factory runs the same craft checks on full manuscripts.