Top AI Document Translation Models
We put them through one blind benchmark, and it settled the ranking — of the models available here, Gemini 3.1 Pro scores highest overall (4.71). For speed or cost-efficiency, a different model wins; see below.
- Overall score 4.71 · Gemini 3.1 Pro
- Median latency 3,598 ms · GPT-5.6 Terra
- Cost per quality point $0.00016 · DeepSeek V4 Pro
Eight document scenarios, three passes each, median taken. Scored blind by GPT-5.4 and Claude Opus 4.7, neither of which competed. The overall score averages five dimensions: faithfulness, fluency, terminology, style and formatting. Latency measured sequentially at concurrency 1. Pricing taken from each vendor's official pricing page. All figures verified on 2026-07-10.
Read the full eight-scenario reviewWhat are you translating?
Pick the closest scenario and we'll tell you which model our benchmark favours.
GPT-5.6 Document Translation
OpenAI released GPT-5.6 on July 9, 2026 as a single generation sold in three capability tiers — Sol, Terra and Luna. We benchmarked all three across 8 document scenarios with dual blind judges: Sol averages 4.70 and beats the identically-priced GPT-5.5 while costing less per call, and Terra is the fastest model in the field. Belin Doc offers Sol and Terra.
Document Translation: Sol scored a perfect 5.00 in five of our eight document scenarios, judged blind.
PDF Translation: Pick Sol for quality or Terra for speed, with layout, tables and terminology preserved.
GPT-5.5 Document Translation
Document Translation: Translates entire contracts, books and papers in one pass with a 1M-token context, keeping terminology consistent end to end.
DeepSeek V4 Document Translation
Document Translation: Advanced reasoning preserves complex layouts, tables, and technical terminology with exceptional accuracy.
PDF Translation: Maintains original formatting while delivering nuanced, context-aware translations across 50+ languages.
GPT-4.1 Document Translation
Document Translation: Superior instruction-following ensures consistent style and terminology across long documents.
PDF Translation: Handles complex academic and legal documents with precise contextual understanding.
GPT-5.2 Document Translation
Document Translation: Excels in professional literature, mastering terminology and style conversion, like a senior translator.
Video Translation: Focuses on creative content localization, intelligently handling cultural metaphors and emotional dubbing.
Gemini 3 Document Translation
Document Translation: Deeply integrated with office software, enabling batch and collaborative translation of multi-format documents.
Video Translation: Supports real-time live stream translation and synchronizes processing of all graphic and text information within the frame.
What people ask before picking a model
The questions worth settling first.
- How were these scores measured?
- Each of the 7 models translated 8 document scenarios, three passes per cell with the median taken — 168 translations in total. GPT-5.4 and Claude Opus 4.7 then scored them blind, neither judge competing, rating faithfulness, fluency, terminology, style and formatting from 1 to 5.
- Should I just pick whichever model scores highest?
- Not necessarily. The top three land within judging noise of each other, so treat them as tied and let cost and latency decide. The overall score only settles the question once the gap is genuinely wide.
- Which model is cheapest?
- By cost per quality point, DeepSeek V4 Pro. That figure divides cost per call by the overall score, so it captures what your money buys in quality — judging by sticker price alone tends to land you on something cheap that scores badly.
- Are these scores user reviews?
- No. They come from blind scoring by judge models, not from users. Belin Doc runs no user rating system, and you will not find star ratings anywhere on our pages or in our structured data.