OpenAI's unreleased GPT-Image-2 spotted on Arena

Arena has quickly become the AI world's favorite staging ground for spotting unreleased tools.

Developer Pieter Levels posted on X that OpenAI's upcoming image model, dubbed GPT-Image-2, has reportedly leaked and is quietly running tests on the Arena platform.

Testing under tape handles

According to Levels, the mysterious model is currently available under three tape-themed aliases: maskingtape-alpha, gaffertape-alpha, and packingtape-alpha.

Early impressions show that the model exhibits remarkably strong world knowledge and significantly improved text rendering capabilities, with Levels noting it may outperform existing options like Nano Banana Pro. OpenAI has yet to officially confirm any details.

Why it matters

If these tape aliases really belong to GPT-Image-2, it takes dead aim at one of image generation's most persistent headaches: legible text. For digital creators and marketers, reliable text rendering inside generated scenes could eliminate hours of manual Photoshop fix-ups.

We think dropping stealth models into blind testing arenas has become the industry's go-to playbook. It lets labs gather real human preference data while letting word-of-mouth hype build organically. While test outputs may still fluctuate before an official release, curious creators should drop by Arena right now to see if they can spot the difference.