According to TestingCatalog, the site that tracks unreleased AI features, OpenAI’s API briefly listed three model names the company has never announced. The first is aurora-green-alpha. Its two siblings, which differ only in the colour, are aurora-blue-alpha and aurora-red-alpha. They surfaced over the weekend. The publication says the names then appeared in blind arena tests, where they generated images, and that the API entries were removed soon afterward.
That is a sighting, not a launch. TestingCatalog’s report cites no statement from OpenAI, and it says no documentation or pricing exists for any of the three. What the public has is a set of strings and a few sample pictures.
TestingCatalog says it learned of the names from an alert on the Dev Mode Discord server and posted them on X on 18 October. It reads the quick removal as an accidental rollout, which is its interpretation. The report does not say whether the models are still reachable in the arena tests.
Two other observers fill in the picture. Lumina Bench, an X account that follows model testing, said all three appear to be image models running on Arena. Intius, another X user, noted that three entries called Aspen, Cypress, and Maple turned up on Design Arena about an hour after the API names. TestingCatalog stresses that nothing ties those three to the Aurora models.
The performance details are thin. In the tests TestingCatalog describes, the red model returned images sooner but with weaker results, and the green model took longer yet produced the nicer-looking pictures. Blue had not been tried at the time of writing. The examples were two cat portraits from green, which the publication called clean and detailed but unremarkable. Those are its own impressions of two images, and no measurement stands behind them.
The leap to a product name is also the publication’s. Lumina Bench speculated this is GPT Image 3, and TestingCatalog says it leans the same way while calling that a guess. Nothing in the three strings says “GPT Image.”
Our reading goes one step further, and it is only a reading. A colour-coded trio appearing together looks like a family being tested in tiers rather than a single model, which is how a vendor prepares a price ladder of speed against quality. TestingCatalog supplies the precedent. ChatGPT Images 2.5 launched with two API models, GPT-Image-2.5 Sunburst for precision and GPT-Image-2.5 Flare for speed. Back in April, LM Arena hosted a trio of Image V2 variants from OpenAI, each given a tape theme. Against that, the three names could just as easily be rival candidates or checkpoints of one model, and with blue untested the top or bottom rung of any ladder is missing.
TestingCatalog notes that a model called gpt-rosalind-discovery appeared on OpenAI’s pricing page last week and remains unannounced, so a quiet listing can precede an announcement by some distance. For developers building on the current image models, the Aurora names are not yet something to plan around: no endpoint, no price, no spec. A pricing-page entry or documentation page would be the first sign that this has moved from a sighting to a product.
Reported by TestingCatalog on 18 October 2026
