Luna-Lisa-Alpha: OpenAI의 차세대 GPT-image 체크포인트, 새로운 지식 컷오프와 함께 유출되다

The tester who pinned mona-lisa-1's knowledge cutoff to May 2025 has a new codename to share: luna-lisa-alpha, the next checkpoint in what looks like the same Open​AI image-model lineage, and this one reportedly clears the exact bar the last one missed — a recent knowledge cutoff, strong text rendering, and "super realistic" output. The claim comes from a single post on X (@chetaslua, August 19), not from Open​AI, and there is no API, no Arena listing, and no announcement attached to it. But it is the first concrete sign of what comes after the anonymous model that took over LMArena's Image Arena earlier this month, and after gpt-image-2, the current shipping flagship that still tops the image leaderboards.

If it holds, the headline isn't the realism — every checkpoint leak in this series has claimed better skin and less "plastic." The headline is the knowledge cutoff. mona-lisa-1's stale training data was its most disqualifying quirk for practical use, and a fresh cutoff is the single fact that would make a successor genuinely worth waiting for. Here is the leak, the context that makes it legible, and what is still unverified as of today.

유출된 내용이 실제로 의미하는 것

The signal is one message from the X account @chetaslua, the same tester whose "Charlie Kirk test" on mona-lisa-1 was picked up by Chinese tech media in early August as evidence of that model's training cutoff. Paraphrased, the post says: a new G​PT-image checkpoint named luna-lisa-alpha is being tested; it is a new checkpoint, the previous one being monalisa; it was put through the same Charlie Kirk test for knowledge cutoff; this one has a recent knowledge cutoff; great text rendering; super realistic. The post includes an image attachment that shows the test output.

Every one of those claims is a single account's report. Nothing has been reproduced independently, Open​AI has not acknowledged the codename, and there is no watermark analysis, no tokenizer fingerprint, and no Arena appearance attached to it — the three forms of evidence that tied mona-lisa-1 to Open​AI in the first place. Treat this as a lead, not a finding.

Why the knowledge cutoff is the tell

The "Charlie Kirk test" is a blunt instrument for measuring when an image model's training data ends: ask it to reason about something that happened after a known date, and see whether the knowledge is there. For mona-lisa-1, the test produced a conspicuously old answer. Repeated date probes made the model write the year as 2025, and asked about the September 2025 death of the conservative commentator Charlie Kirk, it treated the event as not having happened — consistent with a training cutoff around May 2025. Testers also noted the model had no web-search capability in that build, so the stale cutoff was not papered over by live retrieval.

That mattered for a specific reason. A knowledge cutoff in an image model is a proxy for when the model's vision-and-language backbone finished training — and a May 2025 cutoff in a model that appeared in August 2026 implied the checkpoint was largely frozen from an older run, however fresh the image quality seemed. A "recent knowledge cutoff" on luna-lisa-alpha, if real, says the opposite: the checkpoint was trained on a newer world-model, which is the difference between a model that knows what a 2026 smartphone, interface, or celebrity looks like and one that does not. For an image-generation product, that is not trivia — it decides whether the model can render current brands, products, and cultural references correctly.

There is a caveat worth naming: an image model's knowledge cutoff is not the same thing as a chatbot's, and the leaker's test is a heuristic, not a measurement. But as a signal that Open​AI has moved the training window forward, it is the strongest single datapoint in the post.

The checkpoint trail behind it

luna-lisa-alpha does not appear out of nowhere. Its predecessor mona-lisa-1 surfaced on LMArena's Image Arena on the night of August 9–10, 2026, as an anonymous entry with no byline. The evidence tying it to Open​AI was indirect but layered: testers ran its outputs through Open​AI's official image-verification tool and got a hit on the SynthID watermark Open​AI applies through its Go​gle partnership; the model self-identified as G​PT when asked; and its tokenizer matched G​PT-family models. The "-1" suffix read as a staging label, and Open​AI has a documented pattern of testing image checkpoints under codenames before launch — gpt-image-2 itself ran on Arena under the names Maskingtape, Gaffertape, and Packingtape for roughly two weeks before it shipped on April 21, 2026.

What testers actually saw in mona-lisa-1 tracked closely with the new leak's claims, minus the cutoff. Most reviewers converged on a real reduction in the "plastic skin" look — more natural texture, pores, flyaway hair, camera noise — with @WolfRiccardo calling it "a noticeable improvement in realism, especially in the glossy, synthetic skin." Chinese testers reported big gains in 2D/anime and real-person character consistency. But the model still showed noise and blocky artifacts in places, its instruction-following was rated merely average, and — importantly for this leak — it had a habit of adding too much text into images. The verdict from the most measured tester, @synthwavedd, was "a slight improvement over gpt-image-2, not a big jump."

Against that backdrop, the two adjectives in the luna-lisa-alpha leak that matter are not "super realistic" — that was mona-lisa-1's pitch too — but "recent knowledge cutoff" and "great text rendering." If a follow-up checkpoint fixes the two things reviewers dinged mona-lisa-1 for, Open​AI is iterating in exactly the right order.

The bar a successor actually has to clear

Neither mona-lisa-1 nor luna-lisa-alpha is on any leaderboard, so the reference point stays gpt-image-2, the model Open​AI shipped in April and the current #1 on the image arenas. Its standing and its economics are both on the record:

• gpt-image-2 on LMArena — 1,393 text-to-image Elo and 1,467 image-edit Elo, with an Artificial Analysis Image Arena Elo of 1,338 (as of the May 15, 2026 evaluation shown on its OrcaRouter page).

• gpt-image-2 text and resolution — a knowledge cutoff of December 2025, 2K output, flexible aspect ratios, and the "Thinking" reasoning mode Open​AI added at launch.

• gpt-image-2 price — $8.00 input / $30.00 output per million tokens, billed at provider list price with zero markup through OrcaRouter; p50 time-to-first-token of 10.16 seconds.

The practical read: a successor has to beat a model that already does text well, edits better than anything else on the market, and is cheap enough to run at scale. "Super realistic" alone would not move that needle. A fresh knowledge cutoff and materially better in-image text are the two claims that actually would — which is exactly why those are the claims worth watching.

아직 검증되지 않은 것은 무엇인가?

Almost everything, and it is worth being precise about the shape of the uncertainty:

• The Open​AI link. luna-lisa-alpha has none of the evidence mona-lisa-1 had — no watermark hit, no tokenizer match, no Arena provenance. Its lineage is inferred from the leaker's own framing ("last one was monalisa") and the checkpoint-naming theme, not from any technical fingerprint.

• The knowledge cutoff. "Recent" is a direction, not a date. Until someone runs a dated probe set, we do not know whether the cutoff moved from May 2025 to late 2025, to the first half of 2026, or somewhere in between.

• Text rendering and realism. The claims are qualitative and single-source. mona-lisa-1's own "improved realism" reporting later settled into "slight improvement," so early adjectives have a demonstrated tendency to overshoot.

• Whether it becomes a product at all. Checkpoints leak, get cancelled, or ship under different names all the time. A codename ending in "alpha" sounds like an early candidate in a staged series, not a finished release.

이런 누수를 처리하는 방법

The useful mental model is the one that applied to mona-lisa-1: this is a signal that something is coming, not something that is here. luna-lisa-alpha is not callable from any API today, and there is no way to put real traffic on it. The defensible move is to plan for a successor — and to make testing it, when it ships, cost nothing.

That is the case for routing. gpt-image-2 is live on OrcaRouter right now at $8.00 / $30.00 per million tokens — Open​AI's own list price, passed through at zero markup — so any developer evaluating the image roadmap already has a single key, an Open​AI-compatible endpoint, and automatic failover to a working model. When a successor actually lands, the same setup lets you A/B it against gpt-image-2 under real traffic instead of trusting a tweet: if the new checkpoint underperforms, failover catches it; if it wins, you already know the migration cost is zero, because the switch is a model string, not a renegotiation. Vendor price changes on either model are live the same day they happen, since OrcaRouter passes provider list prices through unchanged.

Until then, the concrete thing remains the shipping model. If you are building on image generation today, the decision that matters is not which unreleased codename to wait for — it is whether the current #1, at list price with no markup, already covers your use case.

다음에 볼 콘텐츠

• Whether luna-lisa-alpha shows up on an arena. mona-lisa-1 went from anonymous Arena entries to a leak-driven writeup in days. If a checkpoint named luna-lisa appears in blind tests, the single-source leak becomes a corroborated trail.

• Whether a watermark or tokenizer fingerprint surfaces. That is the evidence that would tie luna-lisa-alpha to Open​AI the way it tied mona-lisa-1 — or fail to.

• The naming series. Open​AI staged gpt-image-2 under tape-themed codenames, then mona-lisa-1, and now luna-lisa-alpha. The "alpha" suffix suggests more candidates may follow before anything ships, so expect the codename list to grow.

• Any dated probe set. The first person to run a real cutoff measurement will turn "recent" into a number, and that number is the one that decides whether this checkpoint is a meaningful step forward.

The honest read: luna-lisa-alpha is one tester's report of the next step in a checkpoint series that has been quietly building toward something since early August. The single most interesting claim in it — a fresh knowledge cutoff — is also the one most in need of independent confirmation. Watch for Arena appearances and a fingerprint, treat the adjectives as marketing until then, and if the successor does ship, the way to find out what it is actually worth is to route a fraction of real traffic to it against the current #1. That is what the checkpoints are for.

자주 묻는 질문

Is luna-lisa-alpha an official Open​AI model?

Not confirmed. The only source is a single leaker's X post, and Open​AI has said nothing. Its reported behavior and the checkpoint lineage suggest Open​AI, but "suggests" is not "confirms" — the SynthID and tokenizer evidence that backed mona-lisa-1 has not been produced for this one.

When will luna-lisa-alpha be released?

There is no date, and no confirmation the checkpoint will ever ship under that name. If the pattern holds — gpt-image-2 ran publicly under codenames for about two weeks before its April 21 launch — a public test would come before any release, but that is pattern-matching, not a schedule.

Can I use luna-lisa-alpha through an API today?

No. It is not on any API or leaderboard. The only current Open​AI image model you can actually call is gpt-image-2, which is live on OrcaRouter at the provider's list price of $8.00 / $30.00 per million tokens with zero markup.