Theo Browne compares his early GPT-6 Astra experience with release benchmarks and competing models. He emphasizes cost per completed task rather than token price alone, but criticizes the limited initial rollout and cautions that demonstrations and aggregate benchmark rankings do not always match ordinary use.
Theo Browne reports major gains in desktop navigation and 3D creation, including detailed Blender scenes and playable games. His aquarium example looks impressive but initially has poor controls, while generated slides and frontend designs fall short of the strongest demonstrations and still need direction.
Theo Browne describes practical software projects including a music player, a personal media client and performance work on his Lakebed service. He reports substantial latency reductions after testing, while distinguishing these iterative results from a polished product delivered without follow-up.
Theo Browne also recounts failures where the agent acknowledged necessary review fixes but stopped instead of implementing or pushing them. He says a later snapshot reduced that behavior without eliminating it, and notes recurring overengineering, review loops and weak video-editing results. His conclusion is enthusiastic about capability but conditional on the task and the reliability users actually observe.
Watch on YouTube
