Summary & Key Takeaways ​

• Simon Willison evaluated GPT-6 Astra's image generation capabilities.
• Astra was compared against GPT-5.6 Sol, Terra, and Luna using pelican-on-bicycle prompts.
• GPT-6 Astra produced significantly better images across all reasoning levels.
• Even Astra's "low" reasoning level outperformed the best GPT-5.6 Sol results.
• Astra appears more cost-effective due to lower token usage, despite higher per-token pricing.
• Input token counts suggest a potential closer relationship between Astra and Luna models.

Our Commentary ​

I'm genuinely impressed by the leap in quality with GPT-6 Astra. The idea that 'low' reasoning on Astra beats 'max' on Sol is wild. It makes me wonder about the underlying architectural changes. And the token count similarity between Astra and Luna? That's a juicy tidbit. We're seeing rapid, tangible improvements here, and it's both exciting and a little dizzying to keep up.