Improved AI models for everyone!

Readlang’s AI features have moved from GPT-4o mini and GPT-4o to OpenAI’s newer GPT-6 and GPT-6.1 models, for every user. Thanks to everyone who gave feedback while some of you were trying GPT-6 in October (thread).

Free accounts now get context-aware translations too!

  • Translations are now context-aware AI translations (GPT-6 Luna) instead of basic word-only ones, so you get the meaning the word has in that sentence.
  • AI explanations now use GPT-6.1 Terra, the same model as Premium. But you only get 10 of these per day for free.
  • If you’d rather have slightly faster translations without context, you can switch context-aware translations off in Settings or in the reader page’s Settings tab.

Premium

  • Translations and explanations use GPT-6.1 Terra, a big step up from GPT-4o mini.
  • Chat replies and Story Bot both use GPT-6.1 Sol for more natural language.

Premium Plus

  • Translations and explanations use GPT-6.1 Sol, the most accurate model tested.

How much better? On ~20,000 real Readlang translations across 126 language pairs, graded by a slower, more expensive AI model:

Model Score
GPT-4o mini (old Premium) 77
GPT-4o (old Premium Plus) 84
GPT-6 Luna (Free) 87
GPT-6.1 Terra (Premium) 94
GPT-6.1 Sol (Premium Plus) 97

The newer models show a particularly big improvement on less widely spoken languages.

On this benchmark, the new Free tier offers better translations than the old Premium Plus tier! There are of course plenty of other benefits of being on the Premium plans, but it’s nice to know that now even on the free tier you are getting good quality context-aware translations.

Explanations you’ve already seen stay the same when you reopen a word, so your flashcards don’t change unexpectedly. New lookups use the new models. If an old explanation is wrong, use the “Report bad explanation” button to regenerate it.

There’s a chance that this change will have unintentional side effects. Please let me know here if you have feedback, especially if anything seems worse than before.

9 Likes

This is working great for me. For some reason, subtle uses of modal particles or highly stylized/idiomatic German in particular would be fairly often mistranslated with the earlier ChatGPT models and I’d need to plug phrases into DeepSeek, but so far I haven’t encountered a single German translation that hasn’t been spot on! And I’m loving it! And I agree with others that the pithiness is probably a net plus: the accuracy and concision really work together well to give one just enough info to stay immersed in the text, which is really the essence of what makes the Readlang workflow great. I may still use Deepseek to explain, in depth, really difficult constructions or subtle jokes or whatever, because its context can be more than a sentence, but I may not have to if the translations are as accurate as I’ve encountered so far - I could just infer all that myself. I’m interested to see how this plays out. In any case a massive boost, thank you Steve!