GPT-6 Astra in Practice: Tests, Use Cases, and Feedback

The author calls it their best work week in a long time, describing Astra as a telescope, compass and propulsion for the Naijm project, which is now moving.

The author compares GPT-6 Astra at launch with the current version using an identical prompt and says today's output is noticeably worse, suspecting OpenAI quietly nerfed it.

The author rounds up the past week's AI news, including GPT-6 Astra's launch, Li Fei-Fei's world model Atlas, and several huge funding rounds, and says the pace is hard to keep up with.

A joint Stanford and Anthropic team released Terminal-Bench 2.0, an 89-task benchmark for agents in real terminal environments, where long-horizon context drift keeps average resolution below 45%.

Users report Astra's quota drains faster, Extra High finishes 30–40% quicker but makes more mistakes, and quality feels worse, leading some to suspect OpenAI cut the compute behind it.

An OpenAI engineer describes a setup where 7 of 8 seats cannot commit changes and only one controlled writer can, plus a context cliff where 271K to 273K input doubles cost and prompt injection success falls to 11.2%.

Given one task, build the most realistic backside possible, Astra worked unattended for 12 hours, improving the scene's graphics, math and physics and redoing anything that looked plastic.

A long-form X article argues that GPT-6 Astra is so capable in sales scenarios that it "should be illegal", and describes how it changes sales workflows.

The author ran Astra xHigh on Devin for 22M tokens at roughly 400k context, using 16% of the weekly quota, and compares Codex's weekly limit calculation.

GPT-6 Astra beat a standard game of Factorio with enemies enabled after nearly 44 in-game hours and 4 days 11 hours on its Codex clock, at an estimated API cost of about $4,500.

The author uses building a character in Three.js as a coding benchmark and says GPT-6 Astra and Fable 5.1 both fail to avoid a PS2-character look.

Given the same Three.js task, Astra acts like a designer who also writes the code, focusing on layout, typography, lighting and animation, while DeepSeek V4.1 Flash concentrates on getting the interaction working.