OpenAI's mysterious image-generation AI is exposed. In actual tests, it can't even compete with Google's Banana - a meme AI model of Google. Netizens: Is this all it can do?

The highlight of this week is about to unfold! The internal code of OpenAI's mysterious image - generation model, GPT Image 2, has leaked, and it will be launched alongside GPT - 5.2. On the other hand, Google's Flash version of the "Banana" and Gemini 3 Flash are also ready.

Missed the release of GPT - 5.2?

Don't worry, OpenAI's "ace in the hole" is on the way!

Just today, two mysterious image models from OpenAI quietly landed on the Design Arena and LM Arena platforms for testing.

Chestnut

Hazelnut

These two models are allegedly GPT Image 2. Some netizens speculate that it will debut this week along with the next - generation flagship model, GPT - 5.2.

Compared with Google's Nano Banana Pro, they possess similar world knowledge, can generate celebrity selfies with comparable image quality, and are particularly good at embedding code in images.

In the following image, it can generate whiteboard writing in a handwritten style (but the logo doesn't seem to be the latest), and can also directly output code when generating images.

Developers who have conducted actual tests have all stated that the image - generation quality of OpenAI still lags behind Google's Nano Banana Pro.

Especially the facial generation effect of characters looks rather fake.

In response, they believe that GPT Image 2 may still be trained based on GP - 4o, but it has improved significantly compared to the previous generation.

The mysterious image - generation model is launched. Can it beat Nano Banana?

So, when OpenAI's Chestnut/Hazelnut goes head - to - head with Google's Nano Banana Pro, which one will prevail?

For this purpose, developer Angel conducted multiple rounds of comparative tests.

Ask Chestnut/Hazelnut to generate an image "showcasing the physical color theory", with a yellow banana and an apple placed one in front of the other, and the overlapping part having a porous, orange skin formed by the mixture of yellow and red.

Obviously, both of OpenAI's models failed.

Left: Chestnut; Right: Hazelnut

This actual test was successfully completed right after the release of Nano Banana Pro, and there were also explorations of different versions.

For another example, ask Hazelnut to create an infographic showing how to make cardamom tea.

Visually, the effects of Hazelnut and Nano Banana Pro are almost the same, and both fully present the production process.

Top: Hazelnut; Bottom: Nano Banana Pro

In the following comparison between Chestnut and Nano Banana Pro, it's obvious that the latter has a higher degree of realism.

Left: Chestnut; Right: Nano Banana Pro

In the image generation of "AGI random thoughts", both Chestnut and Nano Banana Pro have their own highlights.

Left: Chestnut; Right: Nano Banana Pro

When Chestnut and Hazelnut compete simultaneously, the dice generated by Hazelnut have hallucinations (as shown in the following image), and Chestnut performs much better.

Therefore, developers speculate that Chestnut may be the strongest competitor to NBP.

However, some developers' feelings after actual tests are that Chestnut is actually the smaller model.

There are also more actual - test demos:

The Flash version of the "Banana" is also coming

Excitingly and expectantly, Google's new model, "Nano Banana Flash", is also expected to debut this week.

Once Nano Banana was released, there's no need to elaborate on its strength.

It can directly transform the game screen of GTA into a picture with a real - movie texture.

You can edit an image with just one sentence - turn a lemon into a kiwi and a teddy bear into a glass bear.

There are also some amazing