OpenAI launched ChatGPT Images 2.0 on April 21, powered by a new model called GPT-Image-2. On the Arena.ai Text-to-Image leaderboard, it debuted with an Elo score of 1,512. Second place sits at 1,271. That 242-point gap is the largest lead any model has ever held over its nearest competitor in the leaderboard's history.

For context, the entire range from position 4 through position 15 on the same leaderboard spans just 92 points. GPT-Image-2's lead over second place is more than double that entire spread.

The biggest single improvement was text rendering accuracy, which jumped from the 90-95% range to above 99%. Two years ago, asking an image model to generate a restaurant menu produced invented dishes and garbled spelling. This model generates one you could print and put on a table without anyone noticing it was AI-made.

But the more interesting shift is that this isn't really image generation anymore. The model reasons before it draws, searches the web for references, plans layouts in advance, and verifies its own output. In one demo it generated a working QR code embedded in a designed poster from a single prompt.

That's a different category of tool, and it's going to reshape how a lot of teams think about visual content.

Wrote the full breakdown in SavvyMonk newsletter. Do read and subscribe to it, it's free.

https://savvymonk.beehiiv.com/p/openai-just-dropped-the-biggest-image-model-upgrade-in-ai-history