ChatGPT Images 2.0 (openai.com)

Model card for the API endpoint gpt-image-2 (which may or may not reflect the output from ChatGPT Images 2): https://developers.openai.com/api/docs/models/gpt-image-2

API Pricing is mostly unchanged from gpt-image-1.5, the output price is slightly lower: https://developers.openai.com/api/docs/pricing

...buuuuuuuuut the price per image has changed. For a high quality image generation the 1024x1024 price has increased? That doesn't make sense that a 1024x1024 is cheaper than a 1024x1536, so assuming a typo: https://developers.openai.com/api/docs/guides/image-generati...

The submitted page is annoyingly uninformative, but from the livestream it proports the same exact features as Gemini's Nano Banana Pro. I'll run it through my tests once I figure out how to access it.

I know people like to dunk on ChatGPT and Gemini and say Claude is or used to be better, but you can still use worse models when you're out of usage AND make use of Nano Banana and and ChatGPT Image generation with separate limits for your subscription. I think it could make it a more package as a whole for some people (non-programmers). I do like having the option and am excited for which improvements they've done to ChatGPT Image generation because in the past it had this yellow piss filter and 1.5 it sort of fixed it but made things really generic with Nano Banana beating it (altough Gemini also had a too aggressively tuned racial bias which they fixed), it seems the images ChatGPT generates have gotten better.

System card link with safety details https://deploymentsafety.openai.com/chatgpt-images-2-0

direct pdf https://deploymentsafety.openai.com/chatgpt-images-2-0/chatg...

HN submission for a direct link to the product announcement which for some reason is being penalized by the HN algorithm: https://news.ycombinator.com/item?id=47853000

great obfuscation idea - hidden message on a grain of rice

I am hopeful that OpenAI will potentially offer clarity on their loss-leading subscription model. I'd prefer to know the real cost of a token from OpenAI as opposed to praying the venture-funded tokens will always be this cheap.

No gpt-5.5

do they have anything similar to SynthID, or are they just pretending that problem doesn't exist?

I know this is probably mega cherry-picked to look more impressive, but some of the images are terrifyingly realistic. They seem to have put a lot of effort into the lighting.

I caught the last minute of this—was it just ChatGPT Images 2.0?

Price comparison:

GPT Image 2

  Low     : 1024×1024 $0.006 | 1024×1536 $0.005 | 1536×1024 $0.005

  Medium  : 1024×1024 $0.053 | 1024×1536 $0.041 | 1536×1024 $0.041

  High    : 1024×1024 $0.211 | 1024×1536 $0.165 | 1536×1024 $0.165

GPT Image 1

  Low     : 1024×1024 $0.011 | 1024×1536 $0.016 | 1536×1024 $0.016

  Medium  : 1024×1024 $0.042 | 1024×1536 $0.063 | 1536×1024 $0.063

  High    : 1024×1024 $0.167 | 1024×1536 $0.25  | 1536×1024 $0.25

Image generation? Hmm, would be cool if OpenAI also made a video-generation model someday..

I've been trying out the new model like this:

  OPENAI_API_KEY="$(llm keys get openai)" \
    uv run https://tools.simonwillison.net/python/openai_image.py \
    -m gpt-image-2 \
    "Do a where's Waldo style image but it's where is the raccoon holding a ham radio"

Code here: https://github.com/simonw/tools/blob/main/python/openai_imag...

Here's what I got from that prompt. I do not think it included a raccoon holding a ham radio (though the problem with Where's Waldo tests is that I don't have the patience to solve them for sure): https://gist.github.com/simonw/88eecc65698a725d8a9c1c918478a...

Every time a new image gen comes out I keep saying that it won't get better just to be surprised again and again. Some of the examples are incredible (and incredibly scary. I feel like this is truly the point where understanding if something is AI becomes impossible)

[dead]

Can it generate Chibi figures to mask the oligarchy's true intentions on Twitter and make them more relatable?

Suggest renaming this to "OpenAI Livestream: ChatGPT Images 2.0"

The image of the messy desktop with the ASCII art is so impressive - the text renders, the date is consistent, it actually generated ASCII art in "ChatGPT", etc. I was skeptical that it was cherry-picked but was able to generate something very similar and then edit particular parts on the desktop (i.e. fixing content in the browser window and making the ASCII dog "more dog like"). It's honestly astounding, to me at least.

Wow, the difference between AI and non-AI images collapses. I hate the future where I won't be able to tell the difference.

In 5 years and 3 months between DALL-E and Images 2.0 we've managed to progress from exuberant excitement to jaded indifference.

the guys presenting are probably all like 25x smarter than I am but good god, literally 0 on screen presence or personality.

No mention of modifying existing images, which is more important than anything they mentioned.

I think we all know the feeling of getting an image that is ok, but needs a few modifications, and being absolutely unable to get the changes made.

It either keeps coming up with the same image, or gives you a completely new take on the image with fresh problems.

Anyone know if modification of existing images is any better?

Anything better that OpenAI?

This is not as exciting as previous models were, but it is incredibly good. I am starting to think that expressing thoughts in words clearly is probably the most important and general skill of the future.

It's great. Also doesn't seem to have any "slop" standard look, the images it produces are quite diverse.

I would imagine this will hit illustrators / graphics designers / similar people very hard, now that anyone can just generate professional looking graphical content for pennies on the dollar.

One of the images in the blog (https://images.ctfassets.net/kftzwdyauwt9/4d5dizAOajLfAXkGZ7...) is a carbon copy of an image from an article posted Mar 27, 2026 with credits given to an individual: https://www.cornellsun.com/article/2026/03/cornell-accepts-5...

Was this an oversight? Or did their new image generation model generate an image that was essentially a copy of an existing image?