OpenAI's new ChatGPT image model matches Google's Nano Banana Pro on complex prompts

Edits stay more consistent across lighting, composition, and faces

The model also handles image editing differently now. It makes targeted changes without messing up other parts of the image, keeping lighting, composition, and faces more consistent compared to its predecessor. OpenAI says it can handle adding, removing, combining, blending, and transposing image elements.

Use cases include photo editing, virtual try-ons for clothes and hairstyles, and style transformations. OpenAI's demos show things like combining multiple people and a dog from separate photos into one scene, or turning a photo into a movie poster with a golden age Hollywood look.

The model actually follows complex prompts now

The new model is noticeably better at following detailed instructions. In a test with a 6 x 6 grid that needed specific objects in each cell, the new version got the arrangement right, while the old one didn't, OpenAI says. This makes it easier to create images where the placement of elements really matters.

Text rendering is noticeably improved as well. The model can now handle denser, smaller text—so you can get legible snippets of articles, short tables, or infographics with numbers. However, OpenAI admits it still struggles with longer stretches of text, unusual fonts, multiple faces in one image, or producing content in different languages.

We ran our benchmark prompt that requires a detailed, complex, photorealistic scene with an unusual element: a horse riding an astronaut, something the models definitely haven't seen during training. Older models choked on this. But the latest generation, including Flux 2, does much better. The new Image-1.5 performs on par with Google's Nano Banana Pro, and far better than the previous version.

First impressions: the ChatGPT image model produces more intense-looking images compared to Google's Nano Banana Pro. With the same prompt, Nano Banana Pro interprets things more literally and produces a casual photo look rather than a polished photoshoot vibe. That said, this could be a prompting thing.

API prices drop 20 percent despite better performance

Developers can access the model as GPT Image 1.5 through the API. OpenAI says image inputs and outputs are 20 percent cheaper than the previous model. Pricing sits at 8 dollars per million input tokens and 32 dollars per million output tokens for images. Text tokens cost 5 dollars (input) and 10 dollars (output) per million tokens. With the predecessor model GPT-1, images ran between 0.02 cents and 0.19 cents per image depending on quality settings.

OpenAI says the model does a better job preserving brand logos and visual elements, which could matter for marketing and e-commerce use cases. The previous version of ChatGPT image generation is still available as a custom GPT.