🍌 Google released Nano Banana 2.1 - I compared it with gpt-image-2 from OpenAI using the same prompts.
The model is immediately available in the Gemini API, so I ran both through 6 identical tasks: a poster with English text, a cover with Russian text, a street photo, object counting, an infographic, and a meme. On the left in the images is OpenAI (quality high), on the right is Google, both at 1024x1024.
In terms of quality, it's roughly a tie. Both models draw text without a single error, including Cyrillic. Nano Banana is stronger in photorealism: its coffee seller is lively and busy, while OpenAI's is posing for the camera. OpenAI follows the prompt more literally and adds jokes on its own - I didn't ask for the mug "Deployment fueled by caffeine & denial" in the meme.
The main difference is in speed and price:
- gpt-image-2: on average 103 seconds and $0.21 per image
- Nano Banana 2.1: on average 15 seconds and $0.04 per image
Both have the same rate, $30 per million image tokens. The difference is in how many tokens are used per image: OpenAI uses 7024 tokens on high, while Google uses 1120 plus some thinking tokens, because the model thinks first and then draws. OpenAI has a medium mode for $0.05, which I did not test.
https://x.com/GoogleAIStudio/status/2107501303890915550
#google@rvnikita_blog #openai@rvnikita_blog #nanobanana@rvnikita_blog #ai@rvnikita_blog
