The Pelican comparability grid for Astra is fairly attention-grabbing
4th September 2026
I bought entry to GPT-6 Astra this afternoon, so naturally I used it to generate SVGs of pelicans driving bicycles—at low, medium, excessive, xhigh and max reasoning ranges (Astra doesn’t assist reasoning=none). Then I rendered these pelicans in a comparability grid with GPT-5.6 Sol, Terra, and Luna, and past being enjoyable the end result was surprisingly helpful.
See the grid for full high quality pictures. Right here’s the transcript that created the GPT-6 Nova pelicans.
There are just a few attention-grabbing issues that stand out from this grid.
The Astra pelicans are significantly better. The perfect GPT-5.6-Sol pelican (I appreciated xhigh higher than max) continues to be fairly clearly a bunch of summary shapes. Each single one of many Astra pelicans, from low to xhigh, seems to be higher than that. The Astra max one is basically good.
Astra under max nonetheless doesn’t reliably get the pelican legs on each side of the body.
When it comes to price, Astra could also be round twice the value of Sol ($10/million enter, $50/million output, in comparison with $5/$30 for Sol), however it makes use of considerably much less tokens at every of the degrees, making the costs on the completely different ranges nearer than they may in any other case be.
Astra low produces a greater pelican than ANY of the GPT-5.6 Sol fashions at any stage, for 9.55 cents. Spending 10 cents on some other mannequin will get a a lot worse end result.
Have a look at the enter token counts: Astra and Luna each used 16 enter tokens, Sol and Terra used 26. That’s attention-grabbing.
I ponder if Astra and Luna are extra associated to one another than OpenAI let on?
