OpenAI’s ‘mona-lisa-1’ Image Model Spotted on Arena – What the Tests Reveal

Recently, a mysterious image‑generation model surfaced on the Arena platform under the codename mona‑lisa‑1 . Users first noticed it when a prompt asked the model to "generate a picture of your creator", and the model explicitly mentioned OpenAI and Ultraman.

Independent verification using OpenAI’s SynthID detector confirmed that the generated images carried OpenAI’s watermark, leading many observers to conclude the model likely originates from OpenAI, possibly as a successor or test branch of GPT Image 2.

Earlier on August 4, users reported an OpenAI A/B test focusing on prompt‑following ability and text rendering. At that time it was unclear whether the test represented a brand‑new model or a routine update to GPT Image 2. OpenAI has not issued an official statement, but the pattern mirrors previous blind tests of GPT Image 2, which appeared under codenames such as Maskingtape, Gaffertape, and Packingtape before its April release.

After GPT Image 2’s launch, it dominated image‑generation leaderboards with a 242‑point Arena Elo advantage. In direct comparisons, mona‑lisa‑1 shows richer detail and more natural skin texture, addressing the long‑standing "plastic" skin issue of its predecessor. However, some users still observe noticeable noise, blocky artifacts, and only modest overall improvement.

One side‑by‑side test used the same prompt on both models to generate a street portrait. The mona‑lisa‑1 output displayed finer facial freckles, more chaotic hair, and a smoother background blur. Another test asked for a “messy amateur iPhone flash photograph” with an extensive prompt; the result closely resembled a candid phone shot, preserving realistic skin oiliness, localized over‑exposure, and natural sensor noise while avoiding overly polished lighting.

Despite these gains, the model still inserts unwanted text in some images, and its knowledge cutoff appears to be around early 2025. Users who queried the model about the status of public figure Charlie Kirk received an outdated answer, while date‑testing experiments consistently produced the year 2025, suggesting the training data stops near May 2025.

Overall, the community widely believes the "Mona Lisa" codename ties the model to OpenAI, and the only missing piece is an official announcement.