Best AI Model for Manga Translation in 2026: Six Models Tested on 46 Pages
The best AI model for manga translation in 2026: we gave the same 46 manga pages to six models, from GPT-6 to Gemini and Claude Opus, and counted the lines each got wrong.
Serifu · · 7 min read
The best AI models for manga translation in our October 2026 test were gemini-3.8-flash and claude-opus-5-5: neither made a clear mistake on 46 pages. gpt-6-sol made one, gpt-6-luna three, claude-sonnet-5-5 five and claude-haiku-5-5 nine. A bigger model was not always a better translator: a light model beat a middle one.
We make Serifu, a manga translator, and ran this test on 10 October 2026. It was not sponsored by anyone.
| Model | Clear mistakes on 46 pages | Seconds a page | In Serifu |
|---|---|---|---|
| gemini-3.8-flash | 0 | about 15 | Yes |
| claude-opus-5-5 | 0 | 10 to 20, at times over a minute | Yes |
| gpt-6-sol | 1 | about 7 | No |
| gpt-6-luna | 3 | about 8 | Yes, the default |
| claude-sonnet-5-5 | 5 | about 7 | No |
| claude-haiku-5-5 | 9 | about 6 | No |
How the test was run
- The pages. 46 pages and 278 lines: 35 pages of Give My Regards to Black Jack vol. 1 by Shuho Sato (Japanese, free for secondary use), 9 pages of the Korean webtoon 「동백꽃」 by 이호윤 (CC BY), and 2 comic pages in Traditional Chinese. Everything was translated into English.
- Two rounds. First 25 ordinary pages. Then 21 pages picked because translators get them wrong, with what a right translation must do written down before any model ran.
- The same work for every model. Each got the lines' text, the picture of the page, the two pages before it and the same instructions: translate the way an official English release would read. One run per model, nothing edited, each model at its low or no-thinking setting.
- What counts as a mistake. A wrong meaning, the wrong person or sex, a name or term broken, or a word left untranslated. A plainer or livelier voice is not a mistake.
One run on 46 pages shows the mistakes we saw, not a rate. Every picture below is the part of the page beside what each model wrote, word for word.
Who is on the page
Japanese often leaves out he and she. On this page the patient is a woman, and only the picture says so. Both GPT-6 models wrote "him"; the other four read the picture.
One sentence, two balloons
Manga breaks a sentence across balloons, and each balloon needs its own half. gpt-6-luna wrote the whole sentence in both.
Names that must stay the same
第一外科, the First Surgery department, is named twice on one page. claude-sonnet-5-5 called it "General Surgery" the first time and "First Surgery" the second; claude-haiku-5-5 dropped "First" both times.
Doses and drug names
A resident reports heparin at 800 units an hour and a drug called Millisrol. claude-haiku-5-5 lost "an hour" and spelled the drug "Milislon". The other five were right.
Dialect
The Korean webtoon is set in a mountain village and its characters speak Gangwon dialect. Asked "are you working alone?", the boy snaps back: of course alone, would I do it in a crowd? Two Claude models turned the retort around.
The same story has an old insult for someone born a fool. claude-haiku-5-5 took it for a name, "Bennet"; gpt-6-luna wrote "impotent". No model refused the line.
In Japanese, a family from Kyushu thanks a surgeon in their own dialect. All six have the meaning. Three also keep a country voice in English: claude-opus-5-5, gemini-3.8-flash and claude-sonnet-5-5.
Sound effects
ハァ is a sigh. claude-sonnet-5-5 wrote "Pant" for a man who is standing still.
A pun, and a flower that is not what its name says
Neither of these was counted as a mistake for anyone, because there is no single right answer.
A drunk man slurs "that's bad for you" into the word for liver. gpt-6-luna kept both meanings in "It's ba-ad for your li-ver"; gpt-6-sol and claude-opus-5-5 wrote English puns of their own.
The webtoon's title flower, 동백꽃, means camellia in standard Korean. In the story's dialect it is the yellow spicebush, which Korean calls the ginger tree. Five models wrote "yellow camellias". Only gemini-3.8-flash wrote "yellow ginger flowers".
Every mistake counted
- gpt-6-sol (1): the woman patient called "him".
- gpt-6-luna (3): the woman patient called "him"; one sentence written in both balloons; the insult as "impotent".
- claude-sonnet-5-5 (5): "General Surgery" for First Surgery; the dialect retort turned around; a clause the boy never thinks added to his thoughts; ハァ as "Pant"; a blinking light (チカ) as "Click".
- claude-haiku-5-5 (9): the insult as a name; the dialect retort; two lines with who does what reversed; "my surgery department"; the heparin line; a mumble (モゴ) as "sob"; the blinking light left as "Chika"; one line of a doctor's thoughts.
Which of these models Serifu offers
In Serifu you choose the translation model for each project: gpt-6-luna (the default), gemini-3.8-flash or claude-opus-5-5. Whichever you choose, you can change any line afterwards, or translate a page or a line again with another model. See Credits and prices.
Can you just give the pages to ChatGPT, Claude or Gemini?
For the words, yes: these are the same models, and most of their lines in this test were right. Three things decide how well it goes, and a chat window leaves all three to you.
- What the model sees. In this test every model got the picture of the page and the two pages before it, not only the text. Without the picture, nothing says the patient is a woman; without the pages before, names and who is speaking drift.
- The same names all the way through. A volume can be 200 pages. A department, a nickname or a character's name has to come out the same on page 150 as on page 5, which takes a glossary kept for the whole book.
- Getting the words back on the page. A chat gives you text. The page still has to be cleaned and lettered, balloon by balloon.
A manga translator does those three for you. Serifu reads the text on each page, sends it to the model you choose with the page, the two pages before it and your story, characters and glossary, and letters the translation back into its balloons. You can change any line and export the volume as CBZ, PDF or images. See Translate in your browser.
FAQ
Can ChatGPT, Claude or Gemini translate manga?
They translate the text well when they are given the page and the pages before it, as in this test. What they give back is text, not a lettered page: cleaning the original text and setting the translation into the balloons is a separate job, the one a manga translator does.
Which AI model is best for translating manga?
In our October 2026 test of six models on 46 pages, gemini-3.8-flash and claude-opus-5-5 made no clear mistakes. gpt-6-sol made one, gpt-6-luna three, claude-sonnet-5-5 five and claude-haiku-5-5 nine.
Is a bigger AI model always a better translator?
No. claude-sonnet-5-5, the middle model of its family, made five mistakes; gemini-3.8-flash, a light model, made none. The larger claude-opus-5-5 made none either, so size helps in some cases and not in others.
Is Claude or GPT better at Japanese to English manga translation?
It depends on the model, not the family. Claude Opus 5.5 was among the best and Claude Haiku 5.5 the worst; of the two GPT-6 models, sol made one mistake and luna three.
What do AI models get wrong in manga?
In this test: who a line is about when Japanese leaves it out, a sentence split across balloons, a name that must stay the same, drug names and doses, dialect, and sound effects.
Does the model need to see the page, or only the text?
It needs the page. On the page with the woman patient, nothing in the text says she is a woman; the models that got it right read it from the picture.