Ask HN: OpenAI Astra post-launch slump?
Lately it misses obvious details in the code, only to say "oh yeah, I should have mentioned that!"
I explicitly tell it to look at the code in the context, then I start to realize it's making generalizations. "oh yeah, I should have answered from the code!"
The frustrating aspect is that the items listed above are like models from 6-12 months ago. Astra is absolutely amazing and shouldn't perform this way.
Discussion 6 comments
Astra is an implementation model. Use it with instructions like "implement this PRD" or "do that ticket" or "integrate Stripe into this app".
Opus 5.5 does both well so Astra now feels a little worse, but Astra Light has been very capable of completing the job, from what I've experienced in the past few weeks.
One of the hallmarks of Luna is that it's slow. Also, while Luna's price/performance is outstanding (astounding, even) ...it's also sort of dumb. Astra isn't perfect but it's been consistently miles above Luna for me.
I believe you, though. Like many/most in this space, I suspect they mess with this stuff all the time behind the scenes. And they're not necessarily serving you the same thing they're serving me.
Sol is the next biggest model and pretty good, Luna isn't even in the conversation?
I start everything with an Opus 5.5 first step and maybe have 6-Astra review on High effort.