Developer finds Astra only slightly better than 5.6 Sol at 2.5x the price
A developer who used Astra on and off for a few days says it is smarter at finding context and edits code faster, but is not ready for daily use and costs 2.5x 5.6 Sol, so they are sticking with Claude and GPT 5.6.
Everyone is so hyped about GPT 6 Astra... but this was shared with me today It seems to me it's just barely better than 5.6 sol and 2.5x the price. and the thing just isn't that good. I've ran it on and off for a few days and it's a bit more efficient in that it is smarter at finding the context it needs; and it often makes edits faster with less research... but imo it's not ready for prime time. Those 3d model tests i've seen are pretty incredible i'll admit, but i use AI to slop code so for now i'm mostly sticking with claude, glm 5.2 or GPT 5.6 models. I noticed a better improvement from building a local vector embeddings model and language server extension for vs code that deterministically detects SLOP and fixes it and helps any AI model find the necessary context better; and while using that tooling I see token consumption similar to Astra on 5.6 sol for the same tasks. Opinions are my own. https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra