› On launch day for GPT-Image-2.6, I ran the exact same 12 prompts across five flagship image models: OpenAI's new Sunburst, Google's Nano Banana 2.1, Black Forest Labs' FLUX 3, ByteDance's Seedream 5 Pro, and Alibaba's Qwen Image 3 Pro.
› Every test was routed through the exact same API, running each prompt once with no cherry-picking or rerolls.
› Most prompts tested complex typography with diacritics and non-English scripts, which remains the fastest way to expose weak image generators.For a complete breakdown of OpenAI's release, head over to the ChatGPT Images 2.6 overview, or check our generation-over-generation test in 2.6 vs 2.5.Generation speed and API costModelLatency (median)Cost per image (average)GPT-Image-2.6 Sunburst (medium)26 s$0.035GPT-Image-2.6 Flare (medium)8 s$0.035Nano Banana 2.116 s$0.038FLUX 325 s$0.048Qwen Image 3 Pro67 s$0.040Seedream 5 Pro63 s$0.079Pricing is relatively tightly grouped across the board, with the exception of Seedream, which costs more than double the rest.
› Generation speed tells a completely different story: Flare completes a full generation before Qwen or Seedream even start outputting.Rendering complex typography: Polish milk bar menuDiner menu test: the prompt requested Polish menu text.