GPT next-token-prediction models led to the o-series reasoning models.

When will the next paradigm be released?

This is a fuzzy question, but it needs to be a significant advance, and is different in some way. For example, smarter models or larger context windows would not count. Multi-modal mostly doesn't count (e.g. it generates video or accepts audio).

You would need a breakthrough on the order of a near-infinite context window or an autonomous agent API.