OpenAI’s new GPT-5.6 models…Sol, Terra, and Luna…are producing encouraging results.
Yesterday, I used Sol in CotranslatorAI to translate 60,000 words of English to Canadian-French.
I gave the model a robust anchor prompt and plenty of rich context through the Ad-hoc Prompt field.
The total AI token cost was just USD $6.
The punctuation was perfect for the locale (including non-breaking spaces!), and Sol correctly distinguished slogans and program names from the surrounding content. It was also internally consistent. Not much stupid stuff to clean up, which is what allows the translator to post-edit with a focus on delivering premium work.
This was the second batch of a larger job and the second iteration of our GAIT-Augmented MTPE workflow. I used the translator’s helpful feedback on the first 30,000 words to refine the prompt for this second batch.
The larger lesson is not simply that a powerful model can produce good translations.
It’s that with the right human-first approach, you can do even better… professionally better.
Pro tip: I left Reasoning on Auto-set. Increasing it unnecessarily can raise costs without guaranteeing better translations.
One caution: Some users have reported instability, and I experienced a little of that last week. Yesterday, however, Sol handled context windows of 10,000 to 20,000 tokens over and over without a single error.
