OpenAI ships Images 2.6 across ChatGPT, Codex and the API with GIF animation and 50% faster generation
- OpenAI released Images 2.6 in ChatGPT, Codex and the API, claiming generation is 50% faster than before; the company says users now create 4 billion images per week, over half a billion per day.
- Two API models ship: GPT-Image-2.6 Flare for speed, producing an image in as little as six seconds at low quality and improved layouts for posters and product shots, and Sunburst for higher quality and image consistency across edits at medium or high quality.
- Images 2.6 adds GIF output, which OpenAI pitches at animating brand mascots, memes and storyboards; the examples generate looping animations from sprite-sheet prompts and from a hand-drawn owl sketch.
- OpenAI quotes Coco Mao, CEO of OpenArt, saying GIFs are the native language of social media and that GPT-Image-2.6 lets creators produce them from a single prompt.
Hacker News 의견들
Four billion images a week? That's roughly 24 million an hour. Nobody can even see that many, so there has to be a pile of wasted automated effort in there somewhere.
The sprite-sheet grid trick for animation frames is old news, people have been doing it for years. I'm surprised it took an official provider this long to make it explicit. Frame-to-frame adherence is still weak, you can see noise between frames even in their own examples, and SVG is suspicious since that's not how image models work.
Per the docs there's an explicit output_format svg, and it builds a reference raster first, then vectorizes it. Also curious that generation is 50% faster but the price is identical to 2.5.
Price isn't quite identical. 2.6 burns more tokens on low and medium quality, so it works out a bit more expensive in practice.
You also have to pass an extra LLM as svg.model or you get a 400, which isn't documented. I tested Sunburst with gpt-5.5 and gpt-5.6-sol and the SVG results were nothing spectacular, it's about what you already get by asking ChatGPT to convert an existing image.
The webapp never tells you which image model or effort level generated the output, which makes any comparison test useless. That said, one of my harder prompts now passes without cheating.
That looks like a prompting problem to me. Left-right two-tone dye is all over the training data, so meshed grey-blonde is a stretch for 'half blond and grey'.
Still the same gross visible watermark artifacts as the other GPT-Image-2 models. Google's Nano banana watermarking is practically invisible by comparison.
Can it finally do real pixel art? If you mean fixed palette aligned to a grid, no, not without manual post-processing, and the SynthID watermark means postprocessing anyway.
People generating at 1MP makes it worse, that's way higher resolution than traditional pixel art, so the smearing and fringing is obvious to anyone.
What's actually state of the art right now? Nobody even boasts about being the best image generator anymore, I wonder if they're all just good enough.
For pictures I'd still take nano banana. For posters and design work gpt-image wins by a very large margin.
I run a comparative image benchmark and the honest answer is it depends on the metric. Nano Banana Pro and the GPT Image 2 series are the contenders, and OpenAI has led on pure prompt adherence since gpt-image-1 in March 2025, though its fidelity lagged until v2.
2.5 is better than 2.6. I run a prompt-to-design site and pulled 2.6-flare after an hour, it's either too simplistic or produces broken outputs. Sunburst seems solid though, and arena.ai has 2.5 scoring higher.
Opposite experience here, I preferred 2.6 in every example they showed.
Design arena says 2.6 was post-trained to emit less text, which would explain why text elements look more zoomed in and UI designs come out with less detail.