So with GPT-6 Astra, Codex introduced an experimental feature for context management that’s supported to be beneficial for long conversations. Will that experimental feature now also apply to Sol and Luna?
I would also like to point out that it was quite predictable that Terra got discontinued, it didn’t make sense to have it when both Sol and Luna overlapped it.
Lunas insane discount is a game changer, OpenAI knows what they are doing here. Luna at max reasoning effort, even though its not optimal for long conversations, its incredibly intelligent while dirty cheap. Its not even competition anymore.
Whats even crazier is that I’ve underestimated how good Luna actually is. I’ve seen colleges create fantastic things with just Luna medium. This basically means you never have to think about your Codex usage anymore. You can run all day and not
have to worry about your 5h or weekly usage limit. To me, the discounts OpenAI is offering with Sol and Luna is truly a new milestone.
Avoid light and max. Stick to default model selection in Codex. Increase reasoning effort as you go. When the model fails on even xhigh, switch model and start from medium again.
Just thinking loudly here, I guess Co-Author in commits is a hint. PRs where LLMs have been involved is probably another measurement. Github Copoilot has its own account too.
> New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output.
Oh interesting, I can assume what the benefits is for including the Encoder, but whats the downside? I’m thinking GPT (which is decoder only) ruled out Encoder for a reason?
Enc-decs are usually harder to train at frontier scale. Not 100% sure what DeepSeek has done differently here initial read seems to be something related to layer reuse but I just skimmed things so far.
yes, that was surprising to me too. It would be a big deal if they switched to an encoder-decoder model like the original transformer. But I don't think that's what it's doing. One thing is the causal part, so in the original transformer, the encoder was bidirectional, but in this case it is not, so that's one difference. So I think it's an optimization for the prompt/prefill so that the attention is summarized into the output of the encoding layers, rather than all the layers. I just skimmed the paper too so if anyone else has insight, please correct me.
reply