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

A retired engineer says he burns through 17 ChatGPT 20X subscriptions on Astra each week, plus Devin and Claude Max subs, for 30 total, and calls this the rational way to work right now.

The author says he rebuilt DOOM with GPT-6 Astra and will soon share a bloodier version, predicting one person will be able to make AAA games next year.

Boris Power says GPT-6 Astra is a short-term investment to be distilled into smaller models like Luna, and that the biggest problem is users not knowing what AI can already do.

The poster claims Astra 6.1 will arrive at DevDay and beat Opus 5.5, with a stronger model rumored for mid-October, and says Anthropic is finished this year.

In the team's agentic harness, GPT-6 Astra made 10 beds and destroyed 9 End Crystals but ran out of food in the End and died.

The author says the cheaper GPT-6 Sol lacks intelligence, would rather pay more for a model close to Astra's level, and questions why another cheap model is needed when Luna already exists.

A test using GPT-6 Astra in Codex to generate a promo video used 2.4 million tokens, 14% of the weekly limit and 29 minutes, producing unusable off-grid layouts; the author advises against it and suggests switching to Claude.

The author rebuilt their Astra agent so routing, context filling and safety checks run as deterministic code, leaving the model only the hard parts, which meant fewer calls and more predictable failures.

The author used a single prompt to have Astra make a Lemmings game with Polish drivers crashing under a train, and says a playable version will be released tomorrow.

An analysis says Codex still gives more raw quota than Claude, and the real change is pricing: GPT-6 Astra at $10/$50 per million tokens burns through an allowance faster than Opus 5.5 at $4/$20.

Running Mirage's Tesseract on motion-agent tasks, Opus 5.5 cost $19.44 against $158.50 for GPT-6 Astra, and the author argues the metric that matters is cost per approved asset after revisions.

Jacob Posel says he loves both Opus 5.5 and GPT-6 Astra from a couple of weeks ago, and most of all that he can switch between them with no context loss.