Fiction Eval leaderboard · genre
Best AI model for thriller writing
July 2026 — scores recomputed using only the thriller scenes (84 blind matchups, three tries per model). small sample — treat as rough
Grok 4.3 leads, and comes out ahead of #2 in 97% of re-runs. This thriller-only sample is small, so treat the exact order as rough and check where the ranges overlap before reading much into close neighbors.
| # | Model | Thriller score (likely range) |
|---|---|---|
| 1 | Grok 4.3 | 1781 (likely 1716–1836) |
| 2 | Claude Sonnet 4.6 | 1694 (likely 1634–1753) |
| 3 | Claude Opus 4.8 | 1679 (likely 1617–1744) |
| 4 | GLM 5.2 | 1532 (likely 1448–1606) |
| 5 | Gemini 3.1 Pro | 1475 (likely 1393–1551) |
| 6 | DeepSeek V3.2 | 1341 (likely 1270–1408) |
| 7 | Mistral Large | 1341 (likely 1267–1418) |
| 8 | Llama 4 Maverick | 1157 (likely 1106–1235) |
Grok 4.3 on the thriller cold open scene
The radar sweep painted Sector 7 in clean green arcs. Ramirez counted the returns twice, once against the strip board, once against the live feed. All commercial traffic was accounted for—three heavies on the arrival stream, one corporate Lear climbing through eighteen thousand. The only extra paint was a faint primary return at FL310, twenty-eight miles west of waypoint KELSO. No secondary. No data block. Then the transponder lit. Four digits, centered and unambiguous: 7500. Ramirez’s hand moved to the keyboard before the rest of her registered the number. The code was not in the active block list; it hadn’t been issued in the national database since the 2008 revision. She typed the query anyway. No match. No flight plan, no beacon code reservation, no military exercise tag. The aircraft continued straight, no turn, no altitude change, closing the sector boundary at 480 knots. Ninety seconds, she told herself. That was the window before the return crossed into the next center’s airspace or, worse, into the ADIZ buffer. She keyed …
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.