6 ms·

GPT-6 Astra Generally Available

• minimaxir 12d ago
  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.
  • pllbnk 12d ago
    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?
    • minimaxir 12d ago
      Thinking output is already a relatively small part of the total input compared to raw code/text files.
    • vessenes 12d ago
      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.
• rc1 12d ago
  Quite hidden in the mobile app at least. I had to click thinking. The slider goes 5.6 Instant, 5.6 Medium, 5.6 High, 5.6 Extra High, and the surprise 6 Pro
  • minimaxir 12d ago
    I see this as well. Odd there is only 6 Pro.