Deep Dive
Testing at the ceiling
Gostev explains his testing philosophy: medium-difficulty tasks are solved by too many models now, so he builds towards the hardest possible tasks to see where the new model breaks. He landed on 3D visual work because it lets him track capability improvements visually. His benchmark: a voxel representation of historic London spanning medieval, Tudor, and modern eras, all transforming within the same map — a task that requires sustained spatial reasoning across context shifts.
Astra's communication shift
With GPT-5.6, Gostev noticed the model would instantly validate every piece of feedback, reflexively agreeing with criticism in an artificial way. Astra behaves differently — it either pushes back on suggestions or genuinely acknowledges mistakes, mimicking natural human discourse. He calls this shift subtle but important. It's not about technical capability; it's about interaction quality.
The reliability breakthrough
Gostev had been trying to refactor an accumulating codebase — roughly 150,000 lines built up over multiple GPT versions — with GPT-5.6 working but requiring substantial debugging afterward. With Astra, the refactored code ran without intervention. He emphasizes this reduction in hand-holding as the real leap: models are finally mature enough he doesn't need to babysit every step. The constraint now isn't the model; it's his hardware. He moved workloads to a Linux box because CPU became the bottleneck.