OpenAI's new image model, codenamed mona-lisa-1, quietly appeared on the Arena, and after verification by netizens, it was confirmed to have a SynthID watermark, indicating it is from OpenAI. From the test feedback, the realism of the images generated by the new model has significantly improved, with the previously persistent "plastic feel" greatly reduced, and the details in the images are more rich. Its performance in tasks such as creating art paintings and complex infographics is quite impressive.

Interestingly, developers have uncovered many clues. The training of this model may have been completed as early as last year, and its knowledge cutoff date is the same as the existing GPT Image 2, both ending in December 2025. All these signs indicate that it is more like a new checkpoint built on the same foundation. The outside speculation suggests that it is likely a "medium" version of GPT Image 2.5, and more checkpoints may follow. If true, the official version could further push the boundaries of AI-generated images.