Short story: it's smoke and mirrors.

Longer story: This is now how software releases work I guess. Alot is running on open ai's anticipated release of GPT 5. They have to keep promising enormous leaps in capability because everyone else has caught up and there's no more training data. So the next trick is that for their next batch of models they have "solved" various problems that people say you can't solve with LLMs, and they are going to be massively better without needing more data.

But, as someone with insider info, it's all smoke and mirrors.

The model that "solved" structured data is emperically worse at other tasks as a result, and I imagine the solution basically just looks like polling multiple response until the parser validates on the other end (so basically it's a price optimization afaik).

The next large model launching with the new Q* change tomorrow is "approaching agi because it can now reliably count letters" but actually it's still just agents (Q* looks to be just a cost optimization of agents on the backend, that's basically it), because the only way it can count letters is that it invokes agents and tool use to write a python program and feed the text into that. Basically, it is all the things that already exist independently but wrapped up together. Interestingly, they're so confident in this model that they don't run the resulting python themselves. It's still up to you or one of those LLM wrapper companies to execute the likely broken from time to time code to um... checks notes count the number of letters in a sentence.

But, by rearranging what already exists and claiming it solved the fundamental issues, OpenAI can claim exponential progress, terrify investors into blowing more money into the ecosystem, and make true believers lose their mind.

Expect more of this around GPT-5 which they promise "Is so scary they can't release it until after the elections". My guess? It's nothing different, but they have to create a story so that true believers will see it as something different.