[lukax]

> It’s powered by a version of GPT‑5 that’s trained to look across multiple sources to give more comprehensive and accurate answers.

So another GPT-5 fine-tune. Codex also uses a custom GPT-5 fine-tune.

Does fine-tuning make sense now? Or do you have to be OpenAI to fine-tune the models with a mix of existing data and new behaviours?

[Leynos]

I think that you have to be OpenAI (or X, Google or Anthropic) to be able to fine tune models of this scale through reinforcement learning at present.

Look at Tinker for an example of where things might be heading though (https://tinker-docs.thinkingmachines.ai/)

At present though, I get the sense that reinforcement learning at scale is the current battleground (and has been for most of 2025). But we also see over time, the general models adopt the skills taught to the specialized models. Look at how the learning that made codex-1 went into GPT5.

[SilverSlash]

Should we assume "GPT-5" still just means the LLM? It could mean 'GPT-5 the system' which means the model has RAG, tools to use it, and maybe fine-tuned to call those tools.