ChatGPT Images 2.0: OpenAI Revolutionizes Image Generation

⚡

Key Takeaways

1OpenAI has launched ChatGPT Images 2.0, an image generation model that promises a significant advancement over its predecessor.

2Comparative tests with other models, such as Nano Banana 2, show varied performance, but ChatGPT Images 2.0 stands out for its quality.

3The model uses 13,342 tokens to generate complex images, with a cost of about 40 cents per image.

💡Why it matters — This technological advancement from OpenAI could transform the way complex images are generated and used across various sectors.

OpenAI Unveils ChatGPT Images 2.0

OpenAI recently introduced ChatGPT Images 2.0, its latest image generation model, during a live broadcast. Sam Altman, the CEO of OpenAI, compared the evolution of this model to a technological leap similar to that between GPT-3 and GPT-5. This comparison highlights the extent of the improvements brought by this new model.

Initial Tests with gpt-image-1

To assess the progress made, I first tested the old model gpt-image-1 using ChatGPT. The challenge was to find a raccoon in a "Where's Waldo" style image, an exercise that proved frustrating. Even with the assistance of Claude Opus 4.7, which has enhanced image resolution capabilities, the raccoon remained elusive, hidden behind an instruction card in the image.

Comparison with Nano Banana 2 and Pro

I then tested Nano Banana 2 via Gemini. This model successfully identified the raccoon, which was placed quite obviously in the "Amateur Radio Club" booth at the center of the image. The humorous detail "W6HAM" on the booth's sign was also noted. In contrast, Nano Banana Pro, tested in AI Studio, delivered a disappointing performance, failing to spot the raccoon.

Testing the New Model

With these references in mind, I put the new model to the test. I used an updated version of my script openai_image.py, which works with OpenAI's Python client library. Although this library has not yet integrated gpt-image-2, it still allows for its use due to the absence of model ID validation.

Here’s how the test was conducted:

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"

The initial result did not show a raccoon, neither for me nor for Claude. However, after adjusting the output quality and image size settings, I obtained a high-quality image where the raccoon was clearly visible in the bottom left.

Technical Details and Costs

The generated image used 13,342 tokens, charged at $30/million, resulting in a cost of about 40 cents. This pricing makes the model accessible for applications requiring complex images.

Limitations and Perspectives

Despite the advancements, challenges remain. A user on Hacker News, rizaco, attempted to have ChatGPT draw a red circle around the raccoon, but the model failed to solve its own puzzle. This underscores the current limitations of AI models in understanding and interacting with the tasks they generate.

In conclusion, ChatGPT Images 2.0 marks a significant milestone in AI image generation, offering substantial improvements over its predecessors. However, challenges persist regarding the models' ability to solve complex puzzles that they create themselves.