OpenAI Aurora Models Leak: GPT-Image-3 Hints in API

By TechPulse USA Newsroom | 19 October 2026

The most consequential AI launch of the weekend never actually happened — at least, not on the record. Three mystery model names flickered into existence inside OpenAI's own developer API on Saturday, got spotted, screenshotted and argued over, then were quietly scrubbed like they'd never been there. As of Sunday night, OpenAI has said precisely nothing about any of it.

This is how the world's most-watched AI company launches products now: by accident, in public, with no comment. And honestly? It's getting weird.

Here's the straight version of what happened — with every speculation clearly labeled, because on a story like this, the labeling is the journalism.

The OpenAI wordmark. The company's newest image models may already be running on public arenas — it just hasn't told anyone. Image: Wikimedia Commons, public domain.

How three model IDs escaped into the wild

Over the weekend of October 17–18, three unannounced model identifiers surfaced in OpenAI's API: aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha. Then they were pulled.

The first public flag came from @scaling01 — the X account known as Lisan al Gaib, described by Crypto Briefing as an AI researcher — whose post on October 18 sent the IDs ricocheting around AI Twitter. Hours later, TestingCatalog caught the same names through an alert on its Dev Mode Discord server. The outlet's verdict was blunt: the references "were pulled shortly after, which looked like an accidental rollout."

And that's the entire evidentiary foundation. Everything else you've read about these models is scaffolding built on top of one X post and one Discord alert. We re-read all five carrier outlets' full reports ourselves on Sunday, October 18, and cross-checked the major wires; every public claim traces back to that same chain. There is no second independent witness.

Ryan Merket, writing for Runtime Wire, kept the receipts honest: "Lisan al Gaib shared the identifiers on October 18th, but provided no API response or evidence that developers can call them." His warning deserves repeating: "Treating an unverified identifier as a released API model can lead developers to build against access and behavior that have not been documented."

The models that lived just long enough to draw a cat

Here's where it gets genuinely interesting. Within hours of the API sighting, the aurora names started turning up in blind arena image-generation tests — meaning real people were prompting something, somewhere, and getting pictures back.

Lumina Bench, the image-model benchmarking outfit, highlighted that all three appear to be image models currently being tested on Arena. And then came the detail this story will be remembered for: the test outputs were cat portraits. Clean, detailed, unremarkable cat portraits.

TestingCatalog put it better than anyone could: "Of course, two cat portraits won't tell anyone much about quality, but both look clean and detailed. Still, based on the images we've seen so far, none of them look like they'd make you write home about."

The preliminary performance split is the most tantalizing detail we have. In Lumina's testing, red ran faster but with lower quality; green was slower and looked better. The blue model hasn't been tested at all. Keep that two-tier speed-versus-quality pattern in mind — it matters later.

For the record: we can't show you the actual aurora cat portraits. Nobody has published them under a usable license, and we're not in the business of laundering screenshots. So here's a stand-in — a cat portrait generated by DALL-E 3 via ChatGPT 4.0 back in 2024 — to give you the flavor of what leaked arena image tests actually look like.

An AI-generated cat portrait — this one rendered by DALL-E 3 via ChatGPT 4.0 in 2024. The weekend's aurora arena tests produced similarly themed cat images, which we couldn't license; this stands in as illustration. Image: Wikimedia Commons, public domain.

503s, 404s and a wall of silence

Early probers also reported strange error behavior: aurora-red-alpha briefly returned 503 capacity errors — the kind of response you get when something real is sitting behind the endpoint — while green and blue returned 404s. We could not independently verify those error reports; they come from the same observer chain as everything else. But if true, red was alive. Briefly. The others were ghosts.

Let's inventory what OpenAI has actually published about the aurora models. Ready? Nothing. No announcement. No changelog entry. No documentation. No pricing. No stated use case. No capability notes. As Crypto Briefing summarized: "OpenAI has not announced them, documented them, or explained what they do."

Zubiqo Intelligence, in its characteristically dry fashion, put the developer reality in one sentence: "Silent alpha endpoints generate plenty of developer intrigue, but without published pricing, rate limits, or documentation, they remain unusable for any serious production workload."

We looked for an official comment — a post, a forum reply, a spokesperson line — and found nothing. OpenAI's silence is total. That's not unusual for a company that treats its model pipeline like state secrets, but it leaves a colossal gap between what OpenAI knows and what developers are told. Every confident-sounding claim you'll read this week about what aurora "is" should be treated as fan fiction until proven otherwise.

Is this GPT-Image-3? Maybe. That's the honest answer.

Now the speculation — clearly labeled. The strongest theory, advanced by Lumina Bench and endorsed by TestingCatalog, is that aurora is the early scaffolding of GPT-Image-3. TestingCatalog reported that Lumina Bench speculated "this is probably GPT Image 3," adding its own verdict: "we'd lean the same way, though that's still a guess."

It's a guess, but it isn't a crazy one. The circumstantial pile is real. OpenAI just shipped GPT-Image-2.5 in two API flavors — Flare for speed, Sunburst for precision — and a three-color alpha split (red fast-and-rough, green slow-and-pretty) maps neatly onto that speed/precision playbook. Back in April, the company ran three tape-themed Image V2 variants through LM Arena the same way. In August, a "luna-lisa-alpha" checkpoint — following "mona-lisa-1" — leaked through arena testing and was widely read as early GPT-Image-2.5 work. And just last week, OpenAI listed the unannounced "gpt-rosalind-discovery" on its pricing page.

The pattern is unmistakable: OpenAI tests image models under codenames in public arenas, watches the votes roll in, and announces later. Aurora walks like that duck.

But "walks like a duck" is not a species identification. Nobody outside OpenAI has confirmed the GPT-Image-3 link. It remains what TestingCatalog honestly called it — a guess, albeit an educated one. Anyone writing "GPT-Image-3 leaked" as a headline this week is editorializing past the evidence, and the evidence here is thin enough already.

The February 'aurora-alpha' — a red herring worth naming

Complicating everything is an older ghost. An "aurora-alpha" model surfaced on OpenRouter back in February 2026, and unverified theories tied it to a GPT-5.3-Codex-Spark variant — a coding model, not an image model. As Crypto Briefing put it: "That link has not been confirmed by OpenAI. It remains a theory, not a fact."

It is tempting to draw a straight line from February's aurora-alpha to this weekend's color-coded trio. Resist the temptation. The February sighting was on a third-party router, was never confirmed, and pointed at an entirely different model family. Reusing codenames is standard practice at OpenAI; it proves nothing. Treat the February aurora and the October aurora as unrelated until someone shows a connection.

The same discipline applies to one more coincidence. About an hour after the API references appeared, X user Intius spotted three new entries on Design Arena — Aspen, Cypress and Maple — and posted: "Is OpenAI preparing a three-model release soon?" TestingCatalog checked, and the record is clear: nothing ties those names to the Aurora models directly. We're leaving Aspen, Cypress and Maple out of the aurora story entirely, and so should you.

Our take: launches are discovered now, not announced

Step back from the model IDs and look at the machinery. The world's most valuable private AI company cannot keep three strings of text out of its own public API for a full weekend. It cannot stop its "secret" test models from being benchmarked, screenshotted and cat-portraited by the open internet within hours. And when it happens, it says nothing — not a confirmation, not a denial, not even a "we don't comment on unreleased products."

That's not a strategy. It's a botched rollout culture wearing a trench coat of mystique. The would be easier to respect if the walls weren't this porous, or the silence this total. OpenAI wants the launch-day spectacle of a keynote and the deniability of a leak, and increasingly it's getting neither — just a weekend of half-evidence that developers can't build on and journalists can't fully verify.

And let's be blunt about the shoddy state of the evidence itself. We are grading a potential flagship image model on two cat pictures and a Discord screenshot. The 503s and 404s are hearsay. The GPT-Image-3 connection is a guess. The February codename is probably a coincidence. That is the entire story, and it fits in a paragraph — everything else is weather.

Still, the pattern underneath the noise is worth watching, because OpenAI keeps repeating it: codename in the API, arena sightings, scrub, silence, launch. If aurora follows the luna-lisa playbook, expect the real announcement — whatever it turns out to be — within weeks, not months. We'll be watching the API. Somebody has to.

Three things we're tracking, in order: any OpenAI changelog update or documentation that references the aurora identifiers; whether the models appear on third-party routers like OpenRouter, the way February's aurora-alpha did; and whether the GPT-Image-3 theory gets confirmed or killed. Until then, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha remain three entries in a list that no longer exists. OpenAI added the names. The explanation is still pending.

The broader context matters too. This is a company under unusual pressure — already fending off a lawsuit from eight US states demanding a ban on self-improving AI, navigating the fallout from last week's OpenCode breach that exposed developers' API keys, and operating in an industry where, as Demis Hassabis warned AI hacks are "early warning shots" before something bigger. Against that backdrop, a leaky API is the least of OpenAI's problems — but it's the most visible.

Sources

TestingCatalog — "OpenAI testing new 'Aurora' GPT-Image model on Arena" (Oct 18, 2026); Crypto Briefing — "OpenAI quietly lists three aurora alpha models in its API" (Oct 18, 2026); Zubiqo Intelligence — "OpenAI Quietly Lists Three Undocumented Aurora Alpha Models in Developer API" (Oct 18, 2026); Runtime Wire — "Three new Aurora model names pop up in OpenAI's API" (Oct 17, 2026); WinCentral — "OpenAI Aurora Alpha Model IDs Leak: Release Speculation" (Oct 18, 2026). All five carriers were re-read in full during research on Oct 18–19. As of publication, no tier-1 outlet (The Verge, TechCrunch, Reuters, Bloomberg) had covered the story, and OpenAI had issued no official comment anywhere we could find.

TechPulse USA Newsroom covers the AI labs, their leaks and their launches. We verify every fact we print against primary sources; when we can't, we say so.