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

A Japanese poster tested GPT-6 Astra Ultra in Codex and found one-shot output was worse than Opus 5.5 and even the earlier Astra Pro, because it spawned few sub-agents, used far fewer tokens and stopped in under an hour.

The poster says it took six days after launch for GPT-6 Sol's image understanding problem, which also affects computer use, to be discovered, and accuses OpenAI of using Astra internally and offering no compensation.

A developer built Sidewinder, an arcade game with a physical spinning-wheel controller, using Codex, GPT-6 Astra, Blender and WASM, and shared a demo video.

Rather than controlling the game in real time, the author had Astra write code to train a 12k-parameter CNN, observe failures and iterate until it could win the level at 35 fps.

Some users report that GPT-6 Sol High lasts fine on ChatGPT Pro, but switching to Astra burns the quota quickly even on Light, while Opus 5.5 Medium is already enough for daily coding.

A user says GPT-6 Sol in High mode can run all morning on ChatGPT Plus, calling it good value and accurate, but finds GPT-6 Astra unimpressive.

The poster complains that Fable 5, GPT-6 Astra and Opus 5.5 were all weakened shortly after people started using them, and asks what happens between launch day and day 7.

After Astra, Sol, Luna and gpt-image-2.5 shipped in quick succession, the poster wonders how much OpenAI has left in reserve and whether new Codex models are coming.

The author notes OpenAI staff said frontend would be solved by year's end, only three months remain, and Astra has improved but is far from solving it.

Someone reads OpenAI's reply tone as a sign it still has cards to play, guessing there is a new Astra-6.1 model with better performance and usage limits.

A team presents KPI, a promptable kernel for physical interaction, which with their agentic system lets Astra complete long-horizon, contact-rich humanoid tasks zero-shot.

The author notes Astra was quickly overshadowed by Opus 5.5, but still used it with Three.js to build a cave diving game, and suggests watching with sound on.