Something called luna-lisa-alpha has appeared inside Arena’s image-model system, and this time the evidence does not depend on a viral post or a cropped screenshot. I checked the machine-readable model data published with Arena’s own Direct image page. It contains a distinct entry whose public and display name is luna-lisa-alpha.
That establishes one useful fact: the label is genuinely present in Arena’s platform data. It does not establish the much bigger claim now circulating online—that this is a new OpenAI image model.
Here is what Arena’s data actually tells us. Luna Lisa Alpha accepts text input and can also accept images, including multiple image uploads. Its listed output is a square image. The entry has its own internal model ID, is marked as user-selectable, and has no conventional public ranking attached. Yet when I searched the visible model picker during my check on August 20, the name did not appear. That mismatch suggests an experimental, staged, withdrawn, or otherwise limited listing—not a dependable public launch.
Just as importantly, the entry does not name an organization or provider. That omission matters because other records in the same Arena data do identify their makers. The nearby gpt-image-2 entry, for example, explicitly identifies OpenAI as its organization. Luna Lisa Alpha does not.
This also is not simply another name for Mona Lisa 1. Arena’s data contains mona-lisa-1 and luna-lisa-alpha as separate entries with different internal IDs. They share a similar visible capability profile and the obvious “Lisa” naming pattern, which makes a relationship plausible. But a naming resemblance is not a product lineage. Nothing in Arena’s public data says that Luna is Mona’s successor, that both came from the same lab, or that Luna is better.
Why, then, is OpenAI the leading guess? Mainly because earlier discussion around Mona Lisa 1 associated that separate model with GPT Image-like output and alleged OpenAI provenance signals. The repeated “Lisa” naming encourages people to place Luna in the same family. That is an understandable lead for investigation, but it is not a chain of proof. Arena codenames are not commercial product names, output style is not a reliable fingerprint, and evidence attached to Mona cannot simply be transferred to a different Arena ID.
OpenAI’s own documentation makes the present status much clearer. The company’s official model catalog currently lists GPT Image 2 under image generation and editing; it does not list Luna Lisa Alpha. The GPT Image 2 model page calls it OpenAI’s state-of-the-art image model and lists only gpt-image-2 and the dated gpt-image-2-2026-04-21 snapshot. There is no Luna alias, model card, API identifier, pricing page, release note, or announced launch date.
I also searched OpenAI’s official documentation and announcements, Arena’s public help and policy material, its visible image interface, and indexed reporting for the exact codename. I found no official announcement and no independently documented, reproducible evaluation from a major research organization or established technology publication. That absence does not make the Arena entry fake. It does mean performance claims such as “better than Mona Lisa 1” or “the next GPT Image” currently rest on anecdotal samples rather than a disclosed benchmark, stable public access, or vendor documentation.
Provenance tools require equal care. OpenAI’s official provenance-check API can look for known OpenAI signals in an original image, including C2PA and SynthID results. That could become meaningful evidence if an untouched Luna output were tested and the returned record identified an issuer or model. A screenshot, repost, or claim that another “Lisa” model passed such a check is not equivalent. OpenAI also warns that missing signals are inconclusive because metadata or watermarks may be removed or degraded.
Arena itself explains why these mysteries occur. Its published evaluation policy says model providers may preview unreleased systems anonymously: Arena can add an anonymous label, gather votes, share results privately with the provider, and later remove the model. Its Battle Mode guide likewise explains that users compare anonymous outputs before identities are revealed after voting. In other words, a real Arena entry can exist without a public vendor announcement—and without Arena authorizing outsiders to guess the vendor as fact.
So the responsible headline is narrower than the rumor: Arena is testing, or has prepared to test, an unattributed image model called Luna Lisa Alpha. It is separate from Mona Lisa 1. Its public data describes multimodal image-generation capabilities, but not its developer, architecture, quality, price, or release plan.
Could it be an OpenAI candidate? Yes. Is that confirmed? No. Until OpenAI names it, Arena attributes it, or verifiable original files produce model-specific provenance, “the next GPT Image” should remain a hypothesis—not the headline.