Simon Willison tested GPT-6 Astra against GPT-5.6 Sol, Terra, and Luna models by generating SVGs of pelicans riding bicycles at five different reasoning levels. The resulting comparison grid shows that Astra produces significantly clearer images than its competitors, with even the lowest setting outperforming the maximum settings of the older models. While the best Astra output is a recognisable bird, lower settings sometimes fail to place both legs on the frame. The testing also revealed that Astra uses far fewer input tokens than the GPT-5.6 variants, which affects the total cost per request.
This matters because it suggests a shift in how OpenAI prices and structures its new generation of models. Astra charges more per million tokens than Sol, yet the lower token count means the effective price for specific tasks can be competitive or cheaper. The data indicates that the new architecture requires less context to produce high-quality visual results, potentially reducing operational costs for developers. It also hints at a closer technical relationship between Astra and the Luna model, both of which share low input token counts.
* Astra low setting costs 9.55 cents and beats all GPT-5.6 models
* Input token usage for Astra and Luna is 16 versus 26 for Sol and Terra
* Maximum reasoning level on Astra yields the clearest pelican image
