I assume it’s largely a side effect from the final RL in post training?

That’s the step that causes the most significant gains in agentic performance.

But the RL doesn’t care about anything except maximizing the score, so if you only score based on coding benchmarks, anything can happen to the writing style (as long as it doesn’t hurt the coding performance).

That’s why it often gets worse on models that simply had more RL post training from the same base.

Reinforcement learning for specific use-cases like coding that degrade it's writing style... makes sense. Maybe it stands to reason later version of Opus were improved more by this sort of fine-tuning. Feels consistent with the observation of diminishing returns and worsening writing style. Wonder what changed (supposedly) in 5.5.

Does Xiaomis approach help with this? They do all the post training steps at the same time instead of one by one, switch topics after a couple prompts so writing style is mixed with coding and tool use.

Apparently it helps generalize skills between areas, which makes sense when you compare it to how humans learn but I don't know if it's the same for LLMs.

For whatever reason, GPT models simply don't have this problem.

Reasoning models are the same speed. They’re just post trained with RL to do CoT inside tags like before a tag like

There’s no difference in the inference implementation, parameter count, or speed.

There's a difference in the latency distribution between when you submit a query and you see the response, which is what the comment is (clumsily) asking about.

But yeah, there are a lot of factors, so it's hard to answer, and tokens/s isn't the right question.