GPT-Image-1-Mini (simonwillison.net)

OpenAI quietly released gpt-image-1-mini, a smaller, cheaper image-generation model described as “80% less expensive than the large model.” It appeared on the DevDay 2025 announcements page rather than the keynote, and is already usable via the OpenAI API (developers are discovering usage patterns and building small CLI tools to test it). The mini model matters because it provides a dedicated image generator at much lower cost than the large variant, making experimentation and high-volume image generation more practical for researchers, startups, and hobbyists. Key technical details and implications: the API exposes size, output format and quality options (low/medium/high), and reports detailed usage (input_tokens, output_tokens, image_tokens). OpenAI lists image token pricing at $8 per million tokens; model docs suggest low-quality images cost ~$0.005 and medium ~ $0.015. In practice, a default “high” quality run yielded ~4,000–6,000 output tokens (≈ $0.032–$0.048), while a low-quality 1024×1024 JPEG example produced 272 output tokens (≈ $0.002), matching the lower-cost expectations. Generation times observed were ~20–21s on unreliable Wi-Fi. For developers this means predictable, token-based billing, compact API responses with generation metadata, and a viable low-cost path for integrating image synthesis into apps.