OpenAI's Visual Revolution: Everything About GPT Image 2 and Its Surprising Hebrew Support

For years, AI image generation suffered from a glaring "Achilles' heel": the ability to embed readable, accurate text inside the visual. While models like DALL-E 3 or Midjourney created stunning art, the text inside usually looked like gibberish from another world. Now it seems OpenAI has cracked the code with GPT Image 2, the next generation of its built-in image generator, which recently started rolling out to users (and as always with OpenAI, Android users get to taste it first).

This is not just another cosmetic upgrade. It is a deep architectural change that moves from Diffusion models to a "native" multimodal model. What does that mean for us? An unprecedented level of accuracy in understanding the prompt and a writing ability that is no longer limited to English.

The most significant news for the Israeli audience is the leap in the ability to render the Hebrew language. Until now, trying to put Hebrew words in an image ended with reversed letters or unrecognizable marks. In tests of the new model, GPT Image 2 manages to embed complex Hebrew text with remarkable accuracy.

A great example is the ability to take an emotional and complex prompt, such as combining the lyrics of Peer Tasi's song "Ad Acharei HaNetzach" ("Until After Eternity") with a visual dedicated to Memorial Day, and get a result where the words are not only written correctly, but also typographically integrated into the atmosphere of the image. The model understands the cultural context and the linguistic structure of RTL (right to left), which opens the door to creating posts, banners, and marketing content in Hebrew straight from ChatGPT.

According to technical publications and community analyses (such as those on TestingCatalog), GPT Image 2 shows several critical improvements:

• •Understanding complex prompts: The model can follow long, detailed instructions without "forgetting" objects along the way. If you asked for 15 different items in a room, there is a good chance you will find all of them there.
• •Creating user interfaces (UI): The model excels at creating fictional screenshots of apps with clear buttons and text, a powerful work tool for designers and entrepreneurs.
• •Resolution and quality: Built-in support for higher resolutions (up to 4K in some versions) and the ability to produce images in a variety of aspect ratios (like 16:9 for videos) without distorting objects.
• •Speed of execution: Despite the complexity, the model runs more efficiently and faster than its predecessors thanks to the new architecture.

OpenAI is taking a staged rollout approach. For now, the model appears as part of A/B experiments inside the ChatGPT app. Users who got access report a new interface that allows more precise editing of elements inside the image through free conversation.

Unlike DALL-E 3, which was an external model that was "summoned" by ChatGPT, the new model is an integral part of GPT-4o. This lets it use all of the language model's knowledge to understand subtext, quotes from songs, or historical concepts, and turn them into an accurate visual.

We are at a turning point where generative tools are going from nice toys to real work tools. GPT Image 2's ability to handle Hebrew at this level is proof that language barriers are fading. For content creators, business owners, and tech people in Israel, this is an opportunity to produce custom visuals with local text, without having to move to graphic editing software.

If you have Android, now is the time to check whether the update has already reached you. Try giving it prompts with quotes in Hebrew. The results may surprise you.

For more on OpenAI's latest announcements, you can visit the company's official blog.