OpenAI aurora models: what aurora-red-alpha might be

The OpenAI aurora models are three unannounced model IDs, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, that showed up in OpenAI's API model list this weekend. OpenAI hasn't said what they are, but early testers think they're image generation models, likely an unreleased successor to the current ChatGPT image model.

None of this is official. OpenAI hasn't published documentation or pricing, and its changelog doesn't mention the names. What we have is a handful of X posts from people who watch OpenAI's API closely, plus some quick tests.

How the aurora models were spotted

The first post I could find came from @scaling01 at about 3 a.m. UTC on October 18. Lisan al Gaib listed the three names and guessed they might be specialized coding models or faster versions of OpenAI's GPT-6 models.

Fourteen minutes later @AndrewCurran_ posted that "three mysterious new models just showed up in the OpenAI API." His post drew more than 2,000 likes, which is how the names went from API trivia to a weekend topic.

CellCog, a site that tracks model releases, says it called OpenAI's model list with an authenticated account on October 18 and saw all three. Its write-up says they were created on October 17 within about two minutes of each other, which points to a single deployment rather than three separate projects. Other reports say the models were pulled soon after. One roundup says aurora-red-alpha returns capacity errors when called while the other two return "not found," which is what a half-switched-on rollout would look like.

Coding model or image model?

The name has been used before. In February an anonymous model called Aurora Alpha appeared on OpenRouter as a "cloaked" test model. Three days later OpenAI launched GPT-5.3-Codex-Spark, a fast coding model, and many people assumed the two were the same. Neither company ever confirmed it. That's why the first guesses this weekend pointed to coding.

The tests point the other way. @LuminaBench posted that all three appear to be "image gen models" and said they were being tested in Arena, the site where people compare model outputs side by side. TestingCatalog reported that the red version ran faster but produced lower quality images, while green was slower and looked better. Blue hadn't been tested at the time.

@testingcatalog also cooled expectations. It said the references were probably exposed by mistake and then removed, and that regular users likely won't get to try them soon. Its bet is that an update to Daybreak, OpenAI's security program, is the more likely near-term announcement.

One naming note, since search results mix them up: xAI already has an image model called Aurora, which it released in 2024. These OpenAI IDs have nothing to do with it.

My read

I think the image model theory fits best. Three variants, one faster and one better looking, reads like a fast tier and a quality tier. I'd be surprised to see a coding model ship as a color-coded set of three.

Still, everything past "three IDs appeared in the API" is inference from outside OpenAI. The Arena results are anonymous comparisons, the speed and quality notes come from a few testers, and "GPT Image 3" is a guess no one at OpenAI has confirmed. I'd read it as a hint that an image release is close and wait for OpenAI's own spec sheet.

What it means for a business

If you use AI images for social posts or ads, a new OpenAI image model could change your costs and how much cleanup each image needs. For now, leave it alone. Don't wire an unannounced model ID into any tool, because alpha IDs can vanish or change behavior without notice, as these did within hours.

When the model ships with docs and pricing, run your ten most common image requests through it next to what you use now, and compare what each usable image costs once you count the rejects. If you'd like help figuring out which parts of your marketing or admin work are worth handing to AI in the first place, New Face Design's free process audit walks through your week with you and finds the repetitive work worth automating.