A leak named after colors
Over the weekend, three new identifiers briefly appeared in OpenAI's API: aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha. The tip came from an alert on the Dev Mode Discord server and was reported by TestingCatalog, an independent outlet that tracks AI releases ahead of official announcements. The references vanished from the API hours later, which looks like an accidental rollout, but the damage was already done: outsiders had seen the names and went digging.
From the API to the Arena
The next chapter plays out on blind model comparison platforms like LM Arena, where users get results without knowing which system produced them. Independent account LuminaBench says it received aurora-red-alpha and aurora-green-alpha in those tests, both returning cat photos, one closer to a composed portrait and the other more like a casual indoor snapshot. TestingCatalog backed up the finding, saying all three names look like image models being tested at the same time. Separately, a company called CellCog says it saw the three IDs in an authenticated call to OpenAI's model listing, with creation timestamps seconds apart, suggesting they were created together as variants of the same project.
Why this looks like GPT Image 3
OpenAI has a track record of testing multiple versions of the same image model side by side before shipping: that is what happened with GPT Image 2.5, which launched in two variants, Flare for speed and Sunburst for precision, after months running in comparison arenas under temporary names. Red, green and blue fit that same pattern: they are likely not three separate products but three checkpoints or configurations of the same model competing internally before OpenAI decides which one ships. In the few tests seen so far, red came back faster with slightly lower quality, while green took longer and looked better. Nobody has tested blue yet. If the bet pays off, MaxAssistant is watching the selection process for a next generation image generator play out in real time, ahead of any official announcement.
What changes if this holds up
A new OpenAI image generator reopens the direct fight with Google's Nano Banana and with Black Forest Labs' models, right in the week the community expects Google's answer with Nano Banana 3 Pro. For anyone who uses AI to create visual content daily, what is at stake is generation speed, editing precision and detail quality, the same points where GPT Image 2.5 already climbed arena rankings back in September. If OpenAI picks the faster checkpoint as default, the practical gain is cheaper, quicker generation. If it picks the more detailed one, the gain is output quality for people editing images professionally.
What is still unproven
OpenAI has not confirmed anything so far. None of the three names appears on the official API models page, in the API or Codex changelogs, or on OpenRouter's listing. Sam Altman and OpenAI's product communications lead, Tibo Sottiaux, have not commented. All available evidence comes from outside the company: screenshots without full platform context, a third party API call and the reading of historical patterns. This is a signal backed by cross referenced independent sources, not official confirmation, and nobody outside the company has reproduced the test under controlled conditions yet.
Sources
TestingCatalog: OpenAI testing new Aurora GPT Image model on Arena; LuminaBench: Three new GPT image models surface in arena testing; DigIntoAI: Are OpenAI's Aurora Alpha Models Image Generators?; ByteForward: Aurora image models reportedly surface in Arena tests.