Research
We measure things while we build Bublate: which models read and translate comics well, what a chapter costs, and what still needs a person. Each study shows its method and limits, and you can download the raw data and reuse it with credit.
Measured 2026-11-18
Which models translate Korean webtoon dialogue best, and at what cost?
In our November 2026 blind test on 38 Korean webtoon lines, Claude Opus 5.5 scored 8.77 of 10 and DeepSeek V4.1 Flash 1017 scored 8.68 at one 36th of the cost. Seven of the ten models scored between 7.59 and 8.25. Every model scored lower on sound effects than on dialogue. At this size, differences under about 0.3 points are noise.
38 lines, 10 models, 760 scores · 2 CSV files · CC BY 4.0 · updated
Read the study and get the dataMeasured 2026-11-18
Can a model find lettering faults on a translated page?
Partly. In our November 2026 test, 7 models checked 10 translated webtoon pages with 7 known faults, twice each. The best, Gemini 3.8 Flash, found 6 of 7 in both runs for $0.16 to $0.17 a run. The other six found 1 to 5. No cheaper model found the emptied speech bubble reliably, so a person still checks every page.
10 pages, 7 faults, 7 models, 2 runs · 2 CSV files · CC BY 4.0 · updated
Read the study and get the dataMeasured 2026-10-10 to 2026-11-18 (rates checked 2026-11-20)
What does it cost to translate a comic chapter, by hand and by software?
At published rates, translating and lettering a 20-page manga chapter costs $43.20 at the lowest rates we found, $180 at the 2026 averages Anime News Network reported, and $320 at one publisher's stated rates, before sound effects and proofreading. In 8 measured Bublate jobs, a 20-page manga chapter cost $1.906 to run, and Korean webtoon chapters cost $0.476 to $0.890.
8 jobs, 23 published rates · 3 CSV files · CC BY 4.0 · updated
Read the study and get the dataMeasured October 2026 (logged 2026-10-09)
Can a model read brush-drawn Korean sound effects?
Only with the scene. In our October 2026 test of 3 brush-drawn Korean sound effects, read 3 times each, the best model read 6 of 9 from the cropped letters alone. With the scene around the word and a glossary hint, two models read 9 of 9. A larger model read 1 of 9 from the crop alone.
3 sound effects, 3 runs, 9 models · 1 CSV file · CC BY 4.0 · updated
Read the study and get the data
How we test
- Model tests run Bublate's own pipeline code and change one thing at a time.
- When judges score translations, they see the candidates under shuffled labels, so they cannot tell which model wrote which line.
- Every study states its sample size and its limits. A small test shows a direction, not a rate.
- Test chapters are often copyrighted. We publish numbers and descriptions from them, never their art or text.
Reuse the data
All data on these pages is free to reuse under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence: credit Bublate and link to the study. Each study gives a suggested citation. Questions about the data: press@bublate.com. For writers, the press kit has the facts about Bublate.