Lara 3the translation model developed in Rome by Translated, outperforms GPT-5.6 Sol, Claude Fable 5, DeepL and Google Translate in internal tests: 82% of translations were judged accurate without corrections by professional translators. The advantage comes from specialization, not model size.
The Roman company submitted 16,800 translations from English into 21 languages to the blind evaluation of professional translators, who did not know which system had generated each text. The average result was82% for Lara 3versus 69% for OpenAI’s GPT-5.6 Sol, 67% for Anthropic’s Claude Fable 5, 57% for DeepL, and 53% for Google Translate. Lara 3 ranked first in 20 out of 21 language combinations.
The test combined material from the public benchmark WMT25covering books, news and conversations, with proprietary corporate content across the travel, technology and finance industries. Unlike GPT and Claude, Lara 3 is designed exclusively for translation, with a focus on correct terminology, context, consistency between texts and respect for the brand tone.
Learning by Doing
The third generation introduces a training phase “Learning by Doing”: Instead of simply imitating human translations used as a reference, the model generates multiple variations of the same sentence, automatically evaluates them based on the standards of professional reviewers, and refines its output. It therefore repeats the cycle millions of times already in the training phase, even before serving the first customer.
On a performance level, Translated declares a throughput 23 times higher to Claude Fable 5 in single-thread tests, with almost quadrupled translation capacity for the same budget. The declared cost is about a quarter compared to Fable 5, the most expensive model among those compared.
Lara 3 Think
Alongside the general release, which took place on October 16, 2026, October 17 also arrived Lara 3 Thinkthe version with reasoning capacity designed for cases where higher quality is needed. According to Translated it reduces errors by almost half because it takes into account glossaries, style guides and context to ensure greater terminological consistency and respect for the brand voice.
The platform Lara Translate It also integrates translation of images, audio, formatted texts and documents in 72 formats, with a 70% reduction in layout errors. As for costs, the price depends on the length of the source text, not on that of the generated translation. Personalized glossaries, translation memories and specific instructions on tone or terminology incur no extra charges and are included in the base price.
Lara 3 is based on a dataset of 25 million real translations produced by professional translators, the result of the company’s 27 years of experience in the machine translation sector.
