The Pelican comparison grid for Astra is pretty interesting (simonwillison.net)

The recent comparison of the new GPT-6 model, Astra, against its predecessor GPT-5.6 models (Sol, Terra, and Luna) reveals significant advancements in image generation capabilities. Notably, Astra consistently produced higher-quality SVGs of pelicans riding bicycles at various reasoning levels, surpassing even the best outputs of the GPT-5.6 Sol model. While Astra is priced about twice as high per million tokens ($10 for input and $50 for output compared to Sol's $5 and $30), it requires fewer tokens for each output, contributing to a more competitive cost-effectiveness at lower reasoning levels. The results highlight a noteworthy detail: Astra, at its lowest reasoning level, generated superior images compared to all the GPT-5.6 models, making it a compelling option for cost-sensitive tasks. Additionally, both Astra and Luna have similar lower input token counts (16 tokens), compared to Sol and Terra's 26 tokens, sparking speculation about their underlying architectures. This comparison not only showcases Astra's enhanced performance but also raises intriguing questions about potential relationships between different AI models developed by OpenAI, hinting at a deeper evolution in GPT technology.