another possibility is openai uses 1 token for "newline". technically you only need it for flexible image input, but openai may output it anyway to unify both input and output.

For {low, med, high} quality, n = {16, 32, 64} Also because n << n * n, the initial blur image would be very fast to generate, so what people see in ChatGPT (the rolling down visual effect) is likely just animation.

Recently when I look into latest openai image cost, I find this token pattern: square = (n + 1) * n portrait = (n + 1) * (n * 1.5) landscape = (n * 1.5 + 1) * n This may reveal that openai first generate 1 token per row (blur image in ChatGPT), and then row * column tokens.

Introducing JudgeBench – the ultimate benchmark designed to push LLM-based judges to their limits! 🚀 ❓Why do we need a new benchmark for LLM-based judges? As LLMs continues to evolve, their responses become more complex, demanding stronger judges to assess them accurately.