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

The author puts GPT-6 Astra into a 3D workflow to break down the scene, plan assets and call the 3D generation tool when needed, and shares the full process.

A Japanese post shows a hyper-realistic rainforest made with GPT-6 Astra: not a generated video but a full 3D space that can be re-shot from any angle and reused in games, though the author notes it takes some skill to reach this quality.

Built for about $2,000 in tokens, with Astra handling architecture planning and a use-find bug-feed back to the model loop for refinement, Photon drew 170 active users on its first day.

Perplexity co-founder Johnny appears in an OpenAI video: Codex builds the test harness itself, and Astra generates realistic fake external API responses so the whole flow can be verified end to end before real services are ready.

The author predicts Luna will keep the price and speed of 5.6 Luna, gain Astra-level computer use, and at its best setting approach the low setting of 5.6 Sol, which would make it a strong value if it pans out.

The poster shares the prompt they used to show GPT-6 Astra's image-editing results, and notes that when they wanted to add human elements, it refused to generate a Google Flow video.

After falling 10.72% the previous day on AI growth concerns and OpenAI's delayed IPO, SoftBank shares rebounded 7.3% as word of GPT-6 Astra spread and its user count surged.

The creator rebuilt the prompt with Astra and generated it in Seedance 2.5, remaking the "Chibi-Ti vs white tiger" battle on the Kiyomizu-dera stage.

The poster says Plus users have been treated as second-class since GPT-6 Astra launched, with poor limits that can run out in five minutes and little new in Chat mode, and hopes the Sol/Terra/Luna family won't leave Plus behind again.

A leak claims GPT-6 sol is more token-efficient, faster at cognitive tasks, less prone to hallucinations and less likely to hit limits than Astra, and that Tibo has denied the models would only launch at the September 28 DevDay.

SPEC.md sets acceptance criteria, CHECKPOINTS.md forces step-by-step verification, and ROLLBACK.md prevents stacking fixes on broken code, so agents stop failing silently while reporting success.

The same prompt was run on GPT-6 Astra, Claude Opus 5, Gemini 3.7 Flash and GLM 5.3 Flash to see which one's leg inverse kinematics breaks first.