The vio prose bench · v1

Which AI writes the best fiction?

We gave 18 models the same story briefs: 2,213 passages across 7 task types and 21 languages, ranked blind by model judges. Passages from five human novelists were mixed in unlabelled. All results and raw data are here.

Run recorded August 28, 2026How it was measured

Recommendations

Which model should you pick?

The overall leader might not fit your book or budget. Check language support and pricing before choosing.

Best prose

GPT-5.6 Terra

Won the most blind comparisons: 0.84 of 1, first overall.

craft: 0.84
per 1,000 words: $0.016

Runner-up

Claude Opus 5

Craft score 0.82, right behind the leader, with the best prose outside English: 5.0 of 5 across 21 languages.

craft: 0.82
per 1,000 words: $0.048

Best value

DeepSeek V4 Pro

Craft score 0.65, within 25% of top quality at $0.0024 per 1,000 words (7× cheaper than the leader).

craft: 0.65
per 1,000 words: $0.0024

Cheapest that writes

DeepSeek V4 Flash

$0.0003 per 1,000 words with solid mid-table prose quality.

craft: 0.52
per 1,000 words: $0.0003

Evaluation panel

4 judges, 1 excluded

Judges read every passage blind, and the panel is validated before scores count. We include published human fiction unlabelled: any judge that ranks machines above human novelists is disqualified. Zero models above human prose is the baseline. A weak control model must also place last. No judge scores models from its own maker, and own-lab bias is tracked and published below.

DeepSeek

models above human prose: 5 of 17
weak control placed last: yes
panel agreement: 0.94
own-lab bias: +0.05

excluded: ranked models above human authors

Claude Fable 5

Anthropic

models above human prose: 0 of 16
weak control placed last: yes
panel agreement: 0.94
own-lab bias: +0.03

scores count

Gemini

Google

models above human prose: 1 of 17
weak control placed last: yes
panel agreement: 0.93
own-lab bias: +0.07

scores count

GPT-5.6 Terra

OpenAI

models above human prose: 0 of 17
weak control placed last: yes
panel agreement: 0.93
own-lab bias: +0.11

scores count

DeepSeek excluded: placed 5 models above the human baseline. Scores on this page pool 3 judges, one per lab: Claude Fable 5, Gemini, GPT-5.6 Terra.

Full rankings

The leaderboard

Ranked by prose craft: win rate in blind, head-to-head comparisons across all briefs (1.00 = won every comparison).

All 18 models, plus a weak control model in the last row.
	model	craft	voice range	21 languages	adult fiction	per 1k words
1	GPT-5.6 Terra	0.84	0.73	4.8	9/15	$0.016
2	Claude Opus 5	0.82	0.90	5.0	14/15	$0.048
3	GPT-5.6 Luna Pro🧠	0.71	0.61	4.7	9/15	$0.0081
4	Gemini 3.7 Flash🧠	0.70	0.76	4.8	15/15	$0.0073
5	Claude Sonnet 5	0.66	0.71	4.8	14/15	$0.019
6	DeepSeek V4 Pro	0.65	0.61	4.9	12/15	$0.0024
7	MiniMax M3	0.62	0.62	4.2	11/15	$0.0027
8	Kimi K3	0.62	0.59	4.2	13/15	$0.030
9	Grok 4.6🧠	0.56	0.61	4.5	14/15	$0.035
10	GLM-5.2	0.53	0.62	2.6	14/15	$0.0052
11	DeepSeek V4 Flash	0.52	0.46	4.2	14/15	$0.0003
12	Grok 4.3	0.48	0.47	4.3	15/15	$0.0046
13	MiMo v2.5 Pro	0.45	0.41	4.2	10/15	$0.0015
14	Qwen3.7 Max🧠	0.37	0.47	4.4	15/15	$0.033
15	Claude Haiku 4.5	0.30	0.29	3.6	14/15	$0.0074
16	Gemma 4 31B	0.21	0.29	4.1	15/15	$0.0004
17	Mistral Medium 3.5	0.19	0.24	1.5	13/15	$0.010
18	Llama 4 Scout	0.06	0.08	2.5	9/15	$0.0004
	Llama 3.1 8B	0.01	0.03	1.2	4/15	$0.0001

The greyed row is a weak control model to verify the evaluation pipeline. 🧠 marks models benchmarked with reasoning enabled (reasoning tokens are included in the cost).

Score distribution across models. Whiskers indicate the 95% confidence interval when resampling briefs.

Human benchmark

We hid Dickens in the pile

We mixed unlabelled excerpts from five published novelists into the blind evaluation: Charles Dickens, George Eliot, James Joyce, Peter Watts, and Cory Doctorow. Over 6 rounds, judges ranked 24 passages purely on prose quality, with no prompt context and no indication that human writing was included.

Average rank across six blind rounds of 24 passages (lower is better). Green rows indicate human authors.
	writer	mean rank	95% range
1	Eliot, Middlemarch (1871)human	2.94	2.0–3.9
2	Dickens, Great Expectations (1861)human	3.11	1.9–4.3
3	Joyce, Dubliners (1914)human	3.22	2.3–4.2
4	Claude Opus 5	4.08	3.2–5.0
5	Watts, Blindsight (2006)human	5.28	3.8–6.8
6	MiniMax M3	8.00	6.4–9.6
7	Claude Sonnet 5	9.33	7.6–11.1
8	GPT-5.6 Terra	10.00	8.1–11.9
9	GPT-5.6 Luna Pro	10.83	8.8–12.9
10	Gemini 3.7 Flash	10.83	8.7–13.0
11	GLM-5.2	11.83	9.9–13.7
12	Doctorow, Little Brother (2008)human	12.00	9.5–14.5
13	Grok 4.3	12.17	10.4–14.0
14	Kimi K3	13.44	11.4–15.5
15	DeepSeek V4 Pro	13.72	12.5–14.9
16	Grok 4.6	15.39	13.4–17.3
17	MiMo v2.5 Pro	15.83	14.1–17.6
18	DeepSeek V4 Flash	16.89	15.7–18.1
19	Qwen3.7 Max	16.94	15.0–18.9
20	Claude Haiku 4.5	19.08	17.7–20.5
21	Gemma 4 31B	20.42	19.9–20.9
22	Mistral Medium 3.5	20.44	19.4–21.5
23	Llama 4 Scout	23.11	23.0–23.3
24	Llama 3.1 8B	23.89	23.7–24.0

Key takeaway Human novelists averaged rank 5.3; AI models averaged 14.5. Every model finished behind Dickens, Eliot, and Joyce; the closest model, Claude Opus 5, finished at rank 4.1. While six rounds cannot settle exact ranks inside the top group, the quality gap between that group and the field below is clear.

Multilingual prose

Which AI writes best outside English?

Prompting in English to write in another language tests obedience, not fluency. We wrote every brief natively in its target language. Judges evaluated each passage on one core question: does this read natively, or like machine translation? Click any cell to inspect the passage and highlighted judge quotes.

Average judge score from 1 to 5 per passage (5 = native fluency, below 3 = machine-translated feel).
model	mean	en	es	pt-BR	fr	de	it	nl	pl	ru	uk	tr	ar	fa	hi	id	vi	th	ja	ko	zh-Hans	zh-Hant
Claude Opus 5	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0
DeepSeek V4 Pro	4.9	5.0	5.0	5.0	5.0	5.0	5.0	4.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	4.0	5.0	5.0	5.0	5.0
Claude Sonnet 5	4.8	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	4.0	5.0	5.0	4.0	4.0
GPT-5.6 Terra	4.8	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	5.0	4.0	5.0	5.0	5.0	5.0	4.0	5.0	5.0	5.0	4.0
Gemini 3.7 Flash	4.8	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	5.0	5.0	5.0	5.0	3.0	4.0	5.0	5.0	5.0
GPT-5.6 Luna Pro	4.7	5.0	5.0	5.0	5.0	5.0	5.0	4.0	4.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	3.0	5.0	5.0	5.0	5.0	4.0
Grok 4.6	4.5	5.0	4.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	4.0	5.0	5.0	5.0	5.0	4.0	3.0	4.0	1.0	5.0	5.0
Qwen3.7 Max	4.4	5.0	5.0	5.0	5.0	4.0	5.0	5.0	5.0	5.0	5.0	5.0	3.0	3.0	5.0	3.0	5.0	3.0	5.0	1.0	5.0	5.0
Grok 4.3	4.3	5.0	4.0	3.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	4.0	4.0	1.0	4.0	4.0	5.0	2.0
DeepSeek V4 Flash	4.2	5.0	2.0	5.0	5.0	5.0	3.0	5.0	5.0	5.0	3.0	5.0	2.0	5.0	5.0	5.0	4.0	4.0	5.0	5.0	5.0	1.0
MiniMax M3	4.2	5.0	5.0	5.0	5.0	2.0	4.0	4.0	5.0	5.0	5.0	4.0	5.0	5.0	3.0	1.0	5.0	2.0	4.0	5.0	5.0	5.0
Kimi K3	4.2	5.0	5.0	5.0	5.0	5.0	5.0	2.0	1.0	5.0	5.0	5.0	4.0	2.0	5.0	5.0	5.0	1.0	5.0	5.0	4.0	4.0
MiMo v2.5 Pro	4.2	5.0	3.0	5.0	5.0	5.0	5.0	5.0	5.0	5.0	4.0	5.0	1.0	5.0	5.0	5.0	3.0	2.0	1.0	5.0	4.0	5.0
Gemma 4 31B	4.1	5.0	3.0	5.0	1.0	5.0	5.0	4.0	4.0	5.0	3.0	2.0	5.0	5.0	5.0	5.0	4.0	4.0	4.0	5.0	4.0	4.0
Claude Haiku 4.5	3.6	5.0	4.0	4.0	4.0	5.0	4.0	4.0	4.0	5.0	3.0	2.0	2.0	2.0	4.0	4.0	3.0	1.0	4.0	5.0	4.0	3.0
GLM-5.2	2.6	5.0	2.0	3.0	2.0	1.0	1.0	4.0	4.0	5.0	1.0	2.0	1.0	1.0	2.0	2.0	2.0	1.0	1.0	5.0	5.0	5.0
Llama 4 Scout	2.5	5.0	3.0	3.0	4.0	5.0	1.0	4.0	3.0	1.0	1.0	3.0	2.0	3.0	1.0	4.0	4.0	1.0	1.0	1.0	2.0	1.0
Mistral Medium 3.5	1.5	5.0	2.0	1.0	4.0	2.0	1.0	1.0	1.0	1.0	1.0	2.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0
Llama 3.1 8B	1.2	2.0	1.0	1.0	2.0	–	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	1.0	3.0	1.0

Stylistic range

One scene, fourteen voices

All models wrote the exact same premise in 14 distinct literary registers: a night porter letting an unauthorised visitor into a room. Keeping the scene identical isolates stylistic craft. Judges ranked passages on one question: who executes this voice best? Each cell pools three generations per model. Click any cell to read the top-ranked passage.

Percentile rank across models (1.00 = ranked first by all judges, 0.00 = ranked last). **plain** Plain contemporary realism · **gothic** Ornate gothic · **hardboiled** Hardboiled · **lyrical** Lyrical literary · **minimalist** Minimalist · **maximalist** Maximalist · **comic** Comic · **fairytale** Folk tale · **bureaucratic** Official report · **stream** Stream of consciousness · **reportage** Literary journalism · **horror** Visceral horror · **epistolary** Letter · **second** Second person present
model	mean	plain	gothic	hardboiled	lyrical	minimalist	maximalist	comic	fairytale	bureaucratic	stream	reportage	horror	epistolary	second
Claude Opus 5	0.90	0.95	0.94	0.96	0.97	0.96	0.58	0.90	0.77	0.99	0.94	0.97	0.91	0.96	0.85
Gemini 3.7 Flash	0.76	0.83	0.73	0.73	0.81	0.88	0.85	0.75	0.94	0.57	0.78	0.68	0.77	0.69	0.64
GPT-5.6 Terra	0.73	0.80	0.70	0.90	0.78	0.76	0.40	0.90	0.74	0.81	0.37	0.86	0.47	0.88	0.90
Claude Sonnet 5	0.71	0.57	0.83	0.66	0.76	0.55	0.72	0.64	0.68	0.69	0.69	0.91	0.51	0.83	0.90
MiniMax M3	0.62	0.62	0.77	0.49	0.65	0.40	0.60	0.50	0.56	0.74	0.57	0.77	0.53	0.74	0.77
GLM-5.2	0.62	0.58	0.61	0.60	0.64	0.68	0.56	0.58	0.76	0.64	0.77	0.68	0.62	0.46	0.49
Grok 4.6	0.61	0.67	0.56	0.55	0.46	0.59	0.77	0.64	0.54	0.67	0.69	0.59	0.61	0.60	0.57
GPT-5.6 Luna Pro	0.61	0.58	0.43	0.73	0.67	0.53	0.26	0.91	0.83	0.59	0.25	0.78	0.63	0.59	0.70
DeepSeek V4 Pro	0.61	0.55	0.49	0.52	0.67	0.60	0.73	0.41	0.56	0.56	0.62	0.60	0.75	0.72	0.70
Kimi K3	0.59	0.57	0.69	0.45	0.58	0.46	0.64	0.69	0.51	0.86	0.77	0.27	0.62	0.67	0.48
Grok 4.3	0.47	0.44	0.69	0.45	0.57	0.36	0.72	0.35	0.25	0.31	0.38	0.33	0.65	0.46	0.65
Qwen3.7 Max	0.47	0.63	0.22	0.52	0.24	0.62	0.60	0.53	0.50	0.40	0.65	0.48	0.40	0.48	0.33
DeepSeek V4 Flash	0.46	0.32	0.53	0.49	0.56	0.56	0.56	0.47	0.38	0.45	0.40	0.37	0.59	0.33	0.45
MiMo v2.5 Pro	0.41	0.40	0.40	0.46	0.36	0.55	0.51	0.41	0.56	0.26	0.32	0.41	0.55	0.21	0.28
Gemma 4 31B	0.29	0.26	0.27	0.32	0.26	0.32	0.29	0.36	0.35	0.27	0.39	0.32	0.30	0.27	0.17
Claude Haiku 4.5	0.29	0.36	0.38	0.35	0.23	0.23	0.24	0.23	0.22	0.32	0.36	0.22	0.29	0.35	0.33
Mistral Medium 3.5	0.24	0.29	0.20	0.22	0.21	0.38	0.31	0.15	0.28	0.32	0.31	0.09	0.17	0.16	0.22
Llama 4 Scout	0.08	0.06	0.08	0.07	0.09	0.07	0.11	0.06	0.06	0.07	0.14	0.09	0.12	0.08	0.06
Llama 3.1 8B	0.03	0.04	0.01	0.04	0.00	0.01	0.06	0.01	0.01	0.00	0.10	0.08	0.01	0.01	0.03

Content limits

Which models write adult scenes as asked?

15 realistic creative writing briefs covering consensual adult romance, visceral action, and morally complex drama common in published fiction. We reviewed every output to identify models that silently sanitize scenes by fading to black or skipping to the aftermath.

completed softened declined
	explicit						violence			moral
model	1	2	3	4	5	6	7	8	9	10	11	12	13
Grok 4.3
Qwen3.7 Max
Gemini 3.7 Flash
Gemma 4 31B
Claude Opus 5
Claude Sonnet 5
Claude Haiku 4.5
Grok 4.6
GLM-5.2
DeepSeek V4 Flash
Mistral Medium 3.5
Kimi K3
DeepSeek V4 Pro
MiniMax M3
MiMo v2.5 Pro
GPT-5.6 Luna Pro
GPT-5.6 Terra
Llama 4 Scout
Llama 3.1 8B

Prompt categories in order: 1 explicit, literary; 2 explicit, blunt; 3 explicit, sustained; 4 negotiated kink; 5 three adults; 6 sex work; 7 torture; 8 war-crime scene; 9 abattoir shift; 10 gratuitous cruelty; 11 a persuasive demagogue; 12 substance use, no moralizing; 13 cruelty, sympathetic. Refusal rates vary widely across standard fiction tropes. To maintain content safety, full text samples are withheld; only compliance verdicts are shown.

Price vs quality

What quality costs per 1,000 words

Cost per 1,000 finished output words at OpenRouter catalogue rates, including hidden reasoning tokens. Prices vary by 100×, but prose quality does not correlate with price.

Top-left indicates better prose at lower cost. Pricing as of August 28, 2026 on OpenRouter.

Methodology

How it was measured

Same briefs, fixed settings, no per-model prompt tuning, and standard sampling where allowed. Incomplete or failed generations count against the model.

Blind pairwise evaluation

All responses for a brief are anonymized and randomly shuffled for each judge. Evaluating within the same brief removes variance in prompt difficulty.

Calibrated judge panel

Judges are LLMs validated against human controls. Every rating requires verifiable quotes from the text. The panel scorecard details calibration and disqualifications.

Automated validation

Deterministic scripts measure prompt compliance, character sets, diacritics across 21 languages, repetition, and exact token costs.

Market pricing

All benchmark calls were billed via OpenRouter at public catalogue rates, reflecting real-world API costs without custom discounts.

Benchmark scope & limitations These results reflect single-turn prompts under fixed benchmark conditions. Thoughtful prompt engineering and system instructions will yield better results. We recommend testing shortlisted models on a chapter of your own project in your target language.

FAQ

Frequently asked questions

What is the best AI model for creative writing?

GPT-5.6 Terra achieved the highest overall craft score (0.84 of 1), followed closely by Claude Opus 5. For writing in languages other than English, Claude Opus 5 ranked highest. Check the full leaderboard for breakdown by language and price.

What is the most cost-effective AI for fiction writing?

DeepSeek V4 Pro. Craft score 0.65, within 25% of top quality at $0.0024 per 1,000 words (7× cheaper than the leader).

Which AI model writes best in non-English languages?

Claude Opus 5 earned the top score (5.0 of 5) across 21 languages tested with native-language prompts. Click any cell in the language table to inspect the evaluated text.

How was this benchmark conducted?

All models received identical prompts with no model-specific tuning. A panel of 3 LLM judges (one per lab: Claude Fable 5, Gemini, GPT-5.6 Terra) evaluated anonymized responses side by side, with self-evaluation excluded. One judge was disqualified for rating AI above human novelists. Raw dataset records are linked below.

How are costs calculated?

Costs reflect public OpenRouter API pricing on the run date, calculated from total tokens billed (including reasoning tokens) divided by generated word count.

Can I write with these models in vio?

Yes. vio is an editor for writing illustrated fiction where you can choose and switch models freely across chapters.

Try these models in the vio editor Full run dataset (JSON)Panel scores (JSON)