Also is rolling out to the $100/$200 Codex subscriptions, I got access this morning.

Unsurprisingly, it does indeed consume usage at ~2.5x the rate of Sol.

Shouldn't this new reduction in thinking output make the model cheaper to operate? Could it be that the model is the same old LLM, with a bit newer architecture but still doing the same things, including huge amounts of thinking, just that its thinking is less clear to human readers?

	▲	vessenes 21 hours ago \| parent \| next [-]
		Those tend to be under so-called ‘max’ modes, or ‘ultra’ - where you scale up inference time compute and then choose amongst your answers. Astra is new weights.
	▲	minimaxir a day ago \| parent \| prev [-]
		Thinking output is already a relatively small part of the total input compared to raw code/text files.