Inside Praktika's conversational approach to language learning

Praktika, a language-learning startup, is leaning hard on OpenAI's latest models to move past the drill-and-repeat format that's defined the category since Duolingo made owl guilt a business model. Instead of static lesson trees, the company built AI tutors on GPT-4.1 and GPT-5.2 that adjust in real time to how a learner actually talks, hesitates, and messes up mid-sentence.

The pitch is personalization that goes deeper than picking a difficulty level. GPT-4.1 handles a lot of the structured lesson logic — the stuff that needs to be reliable and fast, like grammar correction and vocabulary sequencing. GPT-5.2, being the newer and more capable model, gets used for the harder conversational work: holding context over a longer dialogue, picking up on a learner's specific weak spots, and generating scenarios that feel less scripted than a textbook dialogue about ordering coffee in Madrid.

What's notable here isn't the tech stack itself — plenty of edtech companies are bolting GPT models onto existing products this year — but the emphasis on tracking real-world fluency rather than app-engagement metrics. Praktika frames progress less around streaks and more around whether someone can actually hold a conversation with a stranger, which is a harder thing to measure and a more honest goal than most language apps admit to chasing.

The bigger story is what this says about where consumer AI products are heading. A year ago, most