If anything the chances of converging are lower now, IMO. It used to be 6 months away, in the times of gpt4o / o1, but then scale came into play. Whatever the top labs are doing right now is at a scale that other labs simply can't match. And it seems that RL really benefits from large scale. And as their internal unreleased models improve, so does their training pipelines, diversity of RL environments, filtering, and so on.

While you still need the right architecture (see llama4 w/ bad MoE sparsity), and scale alone won't help you (see the latest Meta models that aren't anything special despite them having scale galore), we also have xAI that got to SotA last year by pure scale. They've since dropped off a bit, but still got within touching distance on pure scale.

The gap becomes obvious once you take them out of distribution w.r.t. the usual benchmarks. Have them work on coding tasks that aren't usually benched for. Have them setup things that aren't "vanilla" setups. They'll do the work, and if you look at the thinking traces they might "mime" the proper reasoning, but the results are sometimes a miss. Looks like a duck, quacks like a duck, but the end result is a chicken :/