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

A short video post in which the author says GPT-6 Astra performs solidly, with a demo clip attached.

People say Sol 6.1 and Astra 6 perform better than at release, suggesting recursive self-improvement; the author plans to switch back from Claude to Codex to check.

A Japanese AI news roundup covering day 7 of Codex's continuous updates, which raised GPT-6.1 Sol's default context from 272k to 500k, plus a test of Astra's quota use on the $500 plan.

The author compares Astra's launch performance with its current state alongside GPT-6.1 Sol and Luna, finding Astra weaker while the newer models got stronger, and wonders whether OpenAI is quietly trying RSI.

A user compares Astra at launch with a week later, saying it improved rather than declined like GPT-6.1 Sol and Luna, and discusses whether OpenAI is heading toward recursive self-improvement.

The author compares launch-day and current videos of Astra, argues it has gotten better rather than worse, cites GPT-6.1 Sol and Luna as supporting evidence, and speculates that OpenAI is doing recursive self-improvement.

The author compares Anthropic's Lean formalization of Fermat's Last Theorem with OpenAI's September and October math announcements, criticizing the latter for declining disclosure quality and vague time definitions.

Using the same prompt of an egg on a forest floor, the author argues Astra only renders the object while Opus 5.5 places it in the scene, and asks readers which side they are on.

A Spanish-speaking developer says Astra's low mode used 30% of a $200 subscription in three tasks at 30 t/s, while Opus 5.5 runs at 80 t/s and lasts longer.

A list of 12 repositories, including Restate, DBOS, Windmill, NATS and PostgreSQL, for building a GPT-6 agent with durable execution across crashes, restarts and disconnected sessions.

After a week of use, Astra fixes bugs at roughly GrokBot level, but it burns through the ChatGPT subscription quota fast and context caching is undocumented and hard to optimize.

The first version took 9 minutes 50 seconds to generate, but repeated revisions were needed to make it feel like a real game, all done on the basic Plus credits.