A mysterious image generation model codenamed "mona-lisa-1" surfaced on the blind testing platform LM Arena, and because its images carry OpenAI's SynthID watermark, it is widely believed to be a new OpenAI image model being secretly tested. The model shows a significant leap over the current flagship GPT Image 2 in visual realism and detail richness, markedly improving the "plasticky" skin texture. When a user asked the model to "generate an image of your creator," it directly named OpenAI and CEO Sam Altman. User @AiBattle_ ran a generated image through OpenAI's verification tool and detected a SynthID digital watermark. OpenAI has not made an official response. Four months ago, before GPT Image 2 was officially released, it underwent similar anonymous testing on Arena under codenames like Maskingtape and Gaffertape, ultimately sweeping image generation benchmarks with a 242-point Arena Elo lead. Developer Riccardo Wolf conducted a side-by-side comparison test, feeding identical prompts into both GPT Image 2 and mona-lisa-1; the street portraits showed mona-lisa-1 delivered more native detail in skin texture, facial freckles, and hair messiness, with more natural background bokeh transitions. In an extreme test targeting the "plasticky" look, testers asked for an image of four people taking a photo with smartphone flash in front of a late-night convenience store, with constraints including "no beauty