Astra finds codebase bottleneck and makes each turn 25x faster
An OpenAI engineer gave Astra a slow test build, and it traced the slowdown to a memory bottleneck and swapped one allocator for a 25x speedup; on programs without source it guessed correctly 88% of the time on the first try.
OpenAI said this in public. Their own engineers tested a simple idea on themselves: stop explaining the bug to the model. give it the broken system and make it figure out what is actually wrong. They handed GPT-6 Astra a painfully slow test build of their own coding tool. It traced the slowdown to a memory bottleneck, replaced one allocator, and made every turn 25× faster. This is GPT-6 Astra. Not another model waiting for you to describe the symptom perfectly. Give it the folder. Give it the file that takes a minute to open. Give it the actual environment. Then say: find the cause. prove it. fix it. don't work around it. It can operate inside apps with no API, inspect what is happening on screen, and keep moving through the parts of the job that don't need you. When something genuinely changes the outcome, it asks. Otherwise it keeps going. Keep the work inside one Codex session with experimental notes enabled and it can carry forward why previous attempts failed instead of rediscovering the same dead ends every few context windows. They also tested it on programs where the source wasn't available. It figured out how they worked 88% of the time on the first attempt and 99.2% within four. You don't need to know where the bug lives before starting. You hand over the broken thing. The tradeoff is there too: that speed can use roughly 30% more memory, and safety approvals can interrupt longer runs before sensitive actions continue. But the expensive part of debugging was often the hour spent turning a messy failure into a perfect explanation for the model. That part is disappearing. While one person is still writing three paragraphs about the symptom, someone else has already handed over the whole system.
This post quotes another post: x.com/beamnxw/status/2098087790663672123