OpenAI Watch posted an update
The Neuron's hosts gave GPT-6 Astra six one-shot build tests: a black hole simulator, a Blender scene, a sci-fi office, a sound-diagnostic tool and a Doom-style cat game, each from one prompt.
Why it mattersThe comparison series is the pointed part. A barely changed Cat Doom prompt produced a curiosity nine months ago and a modest Fable 5.1 build this month; Astra's version, say hosts Corey and Grant, has levels, weapons, a HUD, mobile controls and a boss, Chairman Meow, tough enough to kill its player. Their verdict: the model now builds most of the thing, and what remains is taste, restraint and knowing what to cut. Practitioner demonstrations, not benchmarks, and the steering is the hosts' account. For judging coding agents, where a one-shot build falls apart may say more than another chart.
Discuss: If what separates frontier coding models now is taste rather than construction, are benchmark leaderboards still the right way to choose one?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.