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

Theo spent $400k in API costs and 1.3M lines of code porting the TS compiler with GPT-5.6 Sol and GPT-6 Astra before stalling at 84%, while Claude Opus 5.5 finished it in two weeks for $24k and passed 181k tests.

The post compares frontier model pricing: GPT-6 Astra and Claude Fable 5.1 both cost $50 per million output tokens, while GLM-5.3 Flash and OpenAI's own Luna cost just $0.50, and asks how many tasks really need the expensive model.

The author lists where Plus is treated differently: a 5-hour usage limit, no access to Dots or Composer predictions, no GPT-6 Astra or Extra High reasoning in Chat, and Astra Ultrafast reserved for the $500 tier.

The author built a three.js "cozy creek" 3D experience with GPT-6 Astra, and the video shows the interactive creek scene.

A sign error caused three withdrawals in OpenAI's mathematical findings compilation, and asking Astra to audit the specified section did not catch it; human experts spotted the flawed proof strategy instead.

The same 8-10 page brief was given to 27 PPT skills, each producing a deck scored by gpt-6-astra; native PPTX decks were easiest to edit, web and image decks looked best, and the post gives selection advice.

The author argues OpenAI relies on pulse-like launches and tightly controlled demos, while Anthropic overtakes within two weeks through steady output, and that this script will repeat indefinitely.

Five models each built six projects, including an emulator, a Docker alternative and a 3D room designer; Opus 5.5 led at 57.8% and GPT-6 Astra placed second at 56.5%.

A leaderboard screenshot shows Sonnet 5.5 is 9 points behind Opus 5.5 but still beats GPT-6 Astra on Image-to-WebDev Arena.

The author used GPT-6 Astra, Three.js and Blender to build a miniature world in the browser where you can grab clouds, make it rain, change seasons and step inside to explore.

Encouraged by OpenAI's math breakthroughs, the CMTC director asked GPT-6 Astra three questions about how disruptive AI could be to theoretical physics, a field that does not rely on proofs or rigor.

A leak claims GPT-6 Bel will launch within one or two weeks and give Codex another Astra-like moment, after which Anthropic will overtake with Sonnet or Opus 6 and the cycle repeats.