Access to GPT-6 Astra enabled generation of SVG images of pelicans riding bicycles at multiple reasoning levels. The images were compared to those from GPT-5.6 Sol, Terra, and Luna.
Astra's pelican images, from low to xhigh, outperformed the best GPT-5.6 Sol pelican, which appeared as abstract shapes. Astra's maximum level produced particularly high-quality images, with better leg placement and detail.
Cost analysis shows Astra is roughly twice as expensive as Sol per million tokens, but it uses fewer tokens at each level, making the price difference less significant. Astra low-level images cost about 9.55 cents, outperforming higher-cost models.
Input token counts reveal Astra and Luna both used 16 tokens, whereas Sol and Terra used 26 tokens, suggesting a closer relationship between Astra and Luna.
These findings are relevant for engineers considering model quality versus cost and token efficiency in image generation tasks. Source: https://simonwillison.net/2026/Sep/4/astra-pelicans/