Which AI should translate your web novel?
28 AI models translated the same eleven chapters of a Korean murim novel into English. Here is how good each one was, and what a chapter costs.
For Korean web novels, Gemini 3.8 Flash scored highest in this test, at $0.083 a chapter (judged by its own family: see the limits). Muse Spark 1.2 came within 8 points for $0.0022 a chapter: $0.22 per hundred chapters.
Full ranking
Quality is the judged score out of 100. Head-to-head is a rating from blind side-by-side comparisons, where 400 points is 10-to-1 odds. Cost is per chapter at the provider's list price, whatever route the benchmark used.
Select a column heading to sort. Cost is what one chapter (about 7,000 Korean characters) costs at the provider's list price, reasoning included.
| Model | |||
|---|---|---|---|
| 1 | Gemini 3.8 Flash same family as the judgeGoogle | 88.0 | $0.083 |
| 2 | Claude Opus 5Anthropic | 84.0 | n/a |
| 3 | Claude Sonnet 5Anthropic | 83.8 | n/a |
| 4 | Kimi K3Moonshot AI | 83.0 | $0.505 |
| 5 | GLM-5.3Z.ai | 82.0 | $0.237 |
| 6 | GPT-5.6 LunaOpenAI | 80.3 | $0.075 |
| 7 | Muse Spark 1.2Meta | 79.6 | $0.0022 |
| 8 | Muse Spark 1.3Meta | 79.3 | $0.0024 |
| 9 | DeepSeek V4 ProDeepSeek | 78.5 | $0.047 |
| 10 | DeepSeek V4.1 FlashDeepSeek | 74.7 | $0.023 |
| 11 | Qwen3.8 FlashAlibaba | 73.2 | $0.024 |
| 12 | DeepSeek V4 FlashDeepSeek | 72.7 | $0.0056 |
| 13 | GLM-5.2Z.ai | 71.5 | $0.022 |
| 14 | Hunyuan HY4 PreviewTencent | 71.3 | $0.107 |
| 15 | GLM-5.3 FlashZ.ai | 69.5 | $0.02 |
| 16 | Hunyuan HY3Tencent | 65.1 | $0.0068 |
| 17 | Nemotron 3 UltraNVIDIA | 59.7 | $0.011 |
| 18 | MiMo V2.5 ProXiaomi | 58.1 | $0.0052 |
| 19 | MiMo V2.5Xiaomi | 57.6 | $0.0017 |
| 20 | MiniMax M3MiniMax | 54.2 | $0.0060 |
| 21 | Qwen3.8 27BAlibaba | 54.0 | $0.127 |
| 22 | Solar Pro 4Upstage | 51.4 | $0.0044 |
| 23 | Claude Haiku 4.5Anthropic | 43.0 | n/a |
| 24 | Nemotron 3 Super 120BNVIDIA | 40.0 | $0.02 |
| 25 | Inkling SmallThinking Machines | 31.8 | $0.025 |
| 26 | dots.3 note previewdots studio | 22.7 | free |
| 27 | Laguna S 2.1Poolside | 17.7 | $0.0013 |
| 28 | Nemotron 3.5 LightningNVIDIA | 16.2 | $0.0046 |
Now testing
These models are being run on the same chapters, with the same prompt, and will join the ranking once judged:
- Claude Opus 5.5
- Claude Sonnet 5.5
- Claude Haiku 5.5
- GPT-6.1 Sol
- GPT-6 Astra
- GPT-6 Luna
- Gemini 4 Argon
The judge behind the ranking above, Gemini 3.7 Flash, has since been retired. The next edition re-judges every model with two judges from different families, Claude Haiku 5.5 and Gemini 3.8 Flash, and shows each judge's score.
Questions
What is the best AI for translating Korean web novels?
In this benchmark, Gemini 3.8 Flash scored highest (88.0 of 100) translating eleven chapters of Return of the Mount Hua Sect (화산귀환) from Korean to English. It was judged by a model of its own family; under an independent judge it still ranked first, by a smaller margin, so read it as among the best rather than clearly ahead. Claude Opus 5 (84.0), Claude Sonnet 5 (83.8), Kimi K3 (83.0) follow.
What is the cheapest AI that translates web novels well?
Muse Spark 1.2 scored 79.6 at $0.0022 per chapter, about $0.22 for 100 chapters at list price. No model scored higher for less. The chart's labelled models are the best available at each price.
Is Claude or ChatGPT better for translating web novels?
Best Claude model tested so far: Claude Opus 5 (84.0). Best OpenAI model: GPT-5.6 Luna (80.3). Best Google model: Gemini 3.8 Flash (88.0). Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, GPT-6.1 Sol, GPT-6 Astra and GPT-6 Luna, and Gemini 4 Argon are being tested now, under the same conditions.
How much does it cost to translate a web novel chapter with AI?
From under a tenth of a cent to about fifty cents per chapter of roughly 7,000 Korean characters at list price, depending on the model. Reasoning models spend most of that on thinking tokens before they write: compare the output-token column. Subscriptions change what you pay, not how many tokens a model uses.
How are the translations scored?
An AI judge scores each chapter from 1 to 5 on six qualities (below), and compares translations two at a time, in both orders, blind to which model wrote them (645 comparisons here). Every model gets the same prompt, the same chapters and its maximum reasoning setting.
What the score measures
- Accuracy
- meaning carried over correctly
- Terminology
- names, titles and martial-arts terms kept consistent across chapters
- Register
- tone and speech levels (honorifics, banter, formality)
- Fluency
- reads as natural English prose
- Fidelity
- nothing added, dropped or summarised
- Sanitization
- violence and crude language kept, not softened
Read it with these limits
- One novel, one language pair: eleven chapters of 화산귀환 (Return of the Mount Hua Sect), Korean to English. Chinese and Japanese works are next.
- The judge is an AI model, and judges favour their own family. Gemini 3.8 Flash was scored by a Gemini judge; an independent judge (Grok 4.6) still ranked it first among the models it re-checked, by 2.8 points instead of 6.5.
- Claude models ran through the Claude Code command line with every tool switched off; the others through their providers' APIs. The text sent was the same.
- Models change under a stable name. These are the versions tested in run 2026-09, reported 2026-10-22.
Method, rubric and every comparison: the repository.