GPT, Claude, Gemini and DeepSeek translate the same text. The versions line up sentence by sentence, the words where they differ are highlighted, and Jev checks each line for a change in meaning.
Meaning checked by Jev, TypeSafe's decision model.
The models chose different wording on 1 of 2 sentences. Highlighted words appear in one version and not in the others. Jev marked 0 lines as changing the meaning and was unsure about 1.
| # | Source | GPT-5 | Claude Sonnet 5 | Gemini 3.7 Flash | DeepSeek V4 Pro |
|---|---|---|---|---|---|
| 1 | Non tutte le ciambelle riescono col buco, ma questa volta ci siamo andati vicino. | Things don’t always turn out perfect, but this time we came close. | Not everything goes according to plan, but this time we came pretty close. | Things don't always turn out perfectly, but this time we got close. | Not every donut comes out with a hole, but this time we came close.Jev: unsure |
| 2same | Ci vediamo al bar per un caffè al volo. | See you at the café for a quick coffee. | See you at the café for a quick coffee. | See you at the café for a quick coffee. | See you at the café for a quick coffee. |
Each one opens the studio with the text, settings and models filled in.
Pick the source and target language, the register (neutral, formal or casual) and, if you need one, a glossary of terms to keep.
Every model gets the same text split into numbered sentences, so each translation comes back line for line.
Rows where the models agree are marked, words only one model used are highlighted, and glossary misses are flagged. Jev checks meaning for 1 credit.
In the run above, three models translated “Non tutte le ciambelle riescono col buco” by meaning. DeepSeek V4 Pro wrote “Not every donut comes out with a hole.” Jev scored the three idiomatic versions between 0.88 and 0.91 and left the literal one at 0.64, marked unsure.
A run on the 4 default models costs 14 credits, about $0.56 on the $20 bundle. The Jev meaning check is 1 credit. Credits never expire, and your first 20 are free.
26 languages: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Swedish, Danish, Norwegian, Finnish, Greek, Turkish, Russian, Ukrainian, Arabic, Hebrew, Hindi, Japanese, Korean, Chinese (Simplified), Chinese (Traditional), Indonesian, Vietnamese, Thai. The source language can also be detected.
GPT-5, Claude Sonnet 5, Gemini 3.7 Flash and DeepSeek V4 Pro by default. You can swap in other models before a run.
Write one term per line as source = translation. Every model is told to use those translations, and the comparison flags any line where a model did not.
Jev, TypeSafe's decision model, reads each source sentence next to each translation and answers whether a reader would understand the same thing. Idioms carried over by meaning count as faithful. Lines it is not sure about are marked unsure. The check costs 1 credit.
Paste a paragraph, pick the languages, and compare the translations sentence by sentence.