What Is Luna-Lisa-Alpha? The Arena Checkpoint Everyone Calls GPT Image 2.5
Luna-lisa-alpha is an anonymous image model checkpoint on LMArena that you cannot run. Here is what testers saw, what is confirmed, and what works today.
TL;DR — Luna-lisa-alpha is an unreleased image model checkpoint that started appearing anonymously in blind battles on LMArena. The tester Chetaslua first reported the sighting on X on August 20, 2026. There is no public API, no price, and no model card for this build. You cannot run it anywhere right now, so any website selling access to it is selling you something else.
If you searched for this name, you probably ran into pages promising you can try it right now. Those pages aren't telling you the truth.
There's no public API, there's no model page, and Arena only serves it inside blind battles. What you'll actually find today: a handful of posts, two sample images, and a voting site that pairs it against other models at random.
We went through the dated primary posts, checked the current leaderboard, and tried to test the model ourselves. If you want to run GPT Image 2 in the browser, that model exists and runs today. Here's what's actually happening with this new test build.
What is luna-lisa-alpha?
Luna-lisa-alpha is a checkpoint (a saved training state of a model, not a finished product) that appears anonymously inside LMArena battles.
LMArena is a site where you vote on AI outputs. In a blind battle, two unnamed models answer the same prompt and you pick the better one. Arena only shows you the names after you vote.
A codename in a voting tool is not a launch. It tells you someone is testing an image model. It doesn't tell you a product is out.
The short answers
Here's how the main questions about this checkpoint break down against the checkable facts:
| What people are asking | What is actually known | What it means for you |
|---|---|---|
| Is luna-lisa-alpha a real model? | It appears in LMArena blind battles, reported first on August 20, 2026. | Real enough to see, not real enough to use. |
| Who made it? | Unknown. Testers read it as an OpenAI GPT Image build; OpenAI has said nothing. | Treat "OpenAI" here as a guess with good reasons behind it, not a fact. |
| Can I generate with it? | No. It has no API, no model id, and no product page. | Any site selling access is running a different model. |
| Is it on the leaderboard? | No. The board we read lists 76 models and neither codename is among them. | Nothing about its ranking is knowable yet. |
| When does it launch? | No date exists from anyone. | See the release-date article for the only checkable number. |
| What can I use instead? | GPT Image 2 and Nano Banana 2, which are ranked 1st and 7th on that board, both run on Seadanse. | You can generate today with the models people are comparing it against. |
Leaderboard figures read on August 21, 2026 from the LMArena text-to-image board; the sighting from Chetaslua's August 20 post.
Who saw it, and when
The public trail starts with an independent tester named Chetaslua (@chetaslua). On August 20, 2026, he posted that the codename was being tested on Arena.
That sighting quickly spread across social media. On the same day, Wes Roth on X called it "a mysterious new image model checkpoint" that "has started appearing in public testing", adding that early testers see it as a possible successor to the earlier monalisa checkpoint.
luna-lisa-alpha is being tested , this is new checkpoint last one was monalisa
That first note came from Chetaslua's post on X, which had 131.3 thousand views when we read it on August 21, 2026.
His post quoted his own earlier finding from August 10, 2026. That earlier post documented another anonymous checkpoint called mona-lisa-1.
The one clue worth following
The most interesting detail from the first tester isn't visual style. It's the model's knowledge cutoff (the date after which an AI wasn't trained on any new data).
On August 10, Chetaslua ran his cutoff check on mona-lisa-1. He wrote that the model "think charlie kirk is alive so its older model", which means his test for how recent its data is came back old.
Ten days later, he ran the same test on luna-lisa-alpha. This time, he wrote "this is has recent knowledge cut off".
Keep in mind that this is one person's test, not a benchmark across thousands of prompts. But it's the main clue people cite when they argue luna-lisa-alpha is a newer build.
What testers say changed
After that first post went live, dozens of accounts across English, Chinese, Japanese, Korean, and Turkish communities shared the news.
Almost all of them repeated the same three points from Chetaslua's post:
- A recent knowledge cutoff
- "Great text rendering" inside images
- A "super realistic" visual style
Other accounts added notes about better character consistency and less visible grain. We have to treat those descriptions as one single source repeated many times, not as separate confirmations.
There's only one hard technical detail that anyone published. On August 20, the tester riddi0908 noted that Arena served luna-lisa-alpha outputs at 1672 by 941 pixels, calling it "of the mona-lisa-1 lineage". We confirmed that size on the two sample images attached to Chetaslua's thread.
We can't run the model, so nothing here is our own render of it. What testers posted is the next best thing.
On August 20, 2026, one tester ran a single prompt through both models and posted the two frames together: a station hall at sunset, luna-lisa-alpha first, GPT Image 2 second.
Left: luna-lisa-alpha. Right: GPT Image 2 (medium). One prompt, one run each, posted by @visual58661 on X on August 20, 2026. These frames are the tester's, not ours — nobody can call this checkpoint on demand.
Treat that pair as a sighting, not a benchmark. It's two images, picked by the person who made them.
It's not on the leaderboard, and that matters
If you look at the public rankings, you won't find this model anywhere. We checked the board to see whether either codename had earned a row.
The LMArena text-to-image board on August 21, 2026: 76 models, 5,919,946 votes, and no row for either codename.
The LMArena text-to-image board snapshot from August 10, 2026 includes 5,919,946 votes across 76 models. Neither luna-lisa-alpha nor mona-lisa-1 appears anywhere on that board.
Here's how the top of the official board looks right now:
- Rank 1:
gpt-image-2 (medium)from OpenAI, score 1381 across 70,065 votes. - Rank 7:
gemini-3.1-flash-image (nano-banana-2)from Google, score 1264. - Rank 8:
seedream-5.0-profrom Bytedance, score 1258.
Because luna-lisa-alpha isn't on the board, no one knows its win rate or score. Anyone quoting a benchmark number for it is guessing.
Can you go and see it yourself?
Testers say you can meet the model yourself: open LMArena, switch to Battle Mode, set it to images, and generate. TokenGremlin outlined this method as the way to spot new test runs.
We tried that exact flow on August 21, 2026. We entered a prompt into Battle Mode to see what would load.
Our prompt is still sitting in the box behind the dialog. Battle Mode on August 21, 2026 asked us to sign in before it would generate anything.
Instead of generating images right away, the site asked us to log in or create an account first. You can't hunt for this checkpoint completely anonymously.
More importantly, blind battles pair you with two random models. You can't pick luna-lisa-alpha from a menu or force Arena to test it.
What you can generate today
The top models on the LMArena board are already live. On Seadanse, you can use the GPT Image 2 Generator right inside your web browser.
GPT Image 2 on Seadanse: a prompt box, one optional reference image, and eight aspect-ratio tiles.
When you generate with GPT Image 2 on Seadanse, you get:
- Text prompt generation with no setup
- One optional reference image slot
- Eight aspect-ratio tiles: 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, or auto
- No resolution setting to pick, because this model doesn't have one
If you want to work with several reference images at once, you can also use Nano Banana 2 on the site.
The Nano Banana 2 Realistic AI Image Generator supports 1K, 2K, or 4K outputs and accepts up to 14 reference images at once.
Creating an account is free and needs no card. New signups receive 60 one-time starter credits.
A single still using the GPT Image 2 generator costs 4 credits, so your starter credits give you 15 full generations right away. You can see how credits work on the Seadanse pricing page.
Is it GPT Image 2.5?
Many community posts call this checkpoint "GPT Image 2.5." That name is an unofficial label invented by social media users, not a product name from any company.
OpenAI's published name on the leaderboard is still gpt-image-2. When labs run internal tests, any training checkpoint usually faces one of three paths:
- It finishes testing and ships as a public update.
- It gets folded into another build.
- The team drops the run entirely, and it never launches.
We don't know which path luna-lisa-alpha will take. Calling it a finished release skips steps that haven't happened yet.
What is still unknown
A lot of posts online mix direct sightings with unverified rumors.
On August 20, GodsBoy7777 claimed that luna-lisa-alpha outputs "reportedly trigger OpenAI's verifier." Nobody has shown this first-hand, and OpenAI has said nothing about it.
Here's what remains completely unconfirmed today:
- There is no public model ID.
- There is no pricing sheet.
- There is no official spec sheet.
- There is no confirmed launch date.
If you want a full breakdown of what the odds say about timing, read our guide on the Luna-Lisa-Alpha release date.
The verdict
Luna-lisa-alpha is an unreleased test build that shows up in random blind battles. You can't buy API access to it, and you can't run it on demand anywhere.
If you want to make images today, use the models already sitting at the top of that board.
Written by Nathan Cole. Every post, leaderboard and market page cited here was opened and read on August 21, 2026.
Related articles
- Luna-Lisa-Alpha Release Date: What Is Actually Known (2026) — the only checkable number anyone has on when a new GPT Image model ships.
