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

Across 34,580 verdicts, AI judges picked their own answer 58% of the time versus 34% for humans, with GPT-6 Astra choosing itself 88% of the time and rarely calling a draw.

The poster recounts Altman's talk on GPT-6 Astra, saying most people use only 10% of what ChatGPT can do, and shares the full video and a usage guide.

The author gave the same prompt to Sonnet 5.5 and GPT-6 Astra to build a stretchable 3D jellyfish; Astra's result was glossier but cost $4.29, more than twice Sonnet's $1.95.

Contrasting with Anthropic's statement that it won't release models more capable than Fable, the author calls this buried OpenAI news a big deal.

Bindu Reddy posts that OpenAI still has time, suggesting it cut Astra's price by 50% today.

A post claims GPT-6.1 Astra has been shelved and OpenAI may instead reveal GPT-6o at DevDay, focused on voice and real-time video, though the post is only speculation.

PPTBench reports Astra High at 77.34, but about one in five tasks still fails a gate, showing visual coding agents still have clear gaps.

After the Pro plan's x20 quota was halved, only Sol and Luna followed with price cuts while Astra stayed the same, which the user sees as effectively the same as Anthropic's Fable limits.

Astra High leads PPTBench at 77.34, but semantics, rendering and small-detail reconstruction still commonly fail.

Astra High leads PPTBench with 77.34, but about 19% of tasks fail on semantics, rendered text and visual details.

The author says Astra normally runs at about 30 tok/s, so you can measure the speed to compare when it feels dumber.

The poster says Sonnet 5.5 gives the most usage for the money, followed by Opus 5.5 and then GPT-6 Astra, and guesses OpenAI will launch Astra Minor or GPT-6 Bel at DevDay to compete.