Three model IDs showed up in OpenAI's API without an announcement: aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha. As of the reports we reviewed, OpenAI has not documented, priced or confirmed any of them. This post collects what was reported, separates the observations from the guesses, and explains how to treat a leak like this one.
Everything below comes from third-party reports and social posts. We have not been able to query the API ourselves, so we cannot confirm which IDs are live. Where sources disagree, we say so.
TL;DR: what is known and what is not
| Question | Status |
|---|---|
| Did three aurora IDs appear in the API? | Reported by multiple outlets, first publicly flagged by Andrew Curran on X on October 18, 2026 |
| Has OpenAI confirmed them? | No. Reports say OpenAI's changelog and models page did not mention them |
| When were they created? | One report says between 01:13 and 01:15 UTC on October 17, about 26 hours before anyone noticed |
| Are they still listed? | Reports conflict |
| What are they? | Unknown. An image-model theory has the most detail, but it is unconfirmed |
| Pricing, context window, release date? | None published |
Timeline of the reports
- October 17, 2026, about 01:13 to 01:15 UTC. Per one analysis site, the three model objects were created in this window, which it infers from the API's model metadata. We have not verified this independently.
- October 18, 2026. Andrew Curran posted on X that three mysterious new models had appeared in the OpenAI API, naming all three. The post spread quickly, and several outlets and aggregators wrote it up the same day.
- After the post. Accounts diverge. One says the IDs were removed shortly after. Another says they were still listed at the time of writing. Testers described different responses from each ID: aurora-red-alpha returned capacity-style 503 errors, which can suggest the model is wired up on the backend, while the other two returned 404 not-found errors.
- Documentation check. One report says OpenAI's developer changelog and model page, checked around 11:35 UTC, mentioned no aurora model.
Model listing endpoints can show IDs that are not generally available, so appearing in a list is not the same as being callable, and being callable is not the same as being launched.
What could the aurora models be?
Three theories are circulating. We rank them by how much supporting detail each has, not by how likely we think they are.
| Theory | Basis | Weakness |
|---|---|---|
| Image generation models, possibly a future GPT Image release | TestingCatalog reports that an "Aurora" model appeared on Arena and that all three IDs appear to be image models; red was faster with lower quality, green slower and better looking | Unconfirmed, and the link between the Arena model and the API IDs is inferred |
| Specialized Codex coding models | Speculation from commentators | No reported evidence beyond the "alpha" naming |
| Fast or tiered variants of GPT-6 | Speculation that red, green and blue could be speed tiers | Colors are not an established OpenAI tier naming scheme |
The image-model theory is the only one with outside testing behind it. Even so, an Arena model and an API model ID can be related without being the same thing, and labs often test several candidates under codenames. For background on OpenAI's current image lineup, see our post on ChatGPT Images 2 and gpt-image-2.
Why the red, green and blue names stand out
OpenAI's recent public models use names such as Astra and Sol. Our coverage of those launches is in the GPT-6 Astra launch post and the GPT-6.1 Sol launch post. A three-color scheme with an "alpha" suffix looks more like an internal test series than a product name. In the Arena observations, red behaved like a faster, lower-quality option and green like a slower, better one, which would fit a variant series, but the reports do not describe blue's behavior, and nobody outside OpenAI knows whether the colors map to speed, size, quality or something else.
That pattern, a family of codenamed variants differing along one axis, is common in model testing, but it is an inference, not a fact.
What this means for builders
Do not build on these IDs. Alpha IDs can be renamed, restricted or removed without notice, and an API that returns 404 or 503 for them offers no service guarantee.
Do not scrape the model list for products. If your application discovers models dynamically, make sure it does not route production traffic to unknown or alpha IDs. A short allowlist of documented models avoids surprises.
Do watch the changelog. OpenAI typically documents models when they are ready for developers. If the image-model theory is right, an announcement would likely cover pricing, resolution limits, text rendering and editing features, and you can compare it against your current image workflow then.
Treat benchmark claims from leaked models with caution. Early Arena impressions, such as "red is faster, green is better," are small, informal samples and are easy to over-read. Our guide to how to read AI benchmarks explains why.
How to evaluate a leak like this
- Find the primary source. Here, that is Andrew Curran's post and whatever the API itself returns. Everything else is a summary.
- Separate observation from inference. "The IDs appeared" is an observation. "They are GPT Image 3" is an inference.
- Check for confirmation from the lab. A changelog entry, a blog post or an employee statement counts. Aggregator articles do not.
- Note the timing. One report says the objects existed for about 26 hours before anyone posted, which shows how easily such IDs go unnoticed.
- Wait. Most leaked IDs either launch under a different name or never launch.
How to check the model list yourself
You do not need a leak account to see what your own API key can list. OpenAI's Models endpoint returns the IDs visible to your organization, and the official SDK wraps it:
from openai import OpenAI
client = OpenAI() # reads OPENAI_API_KEY from the environment
ids = sorted(m.id for m in client.models.list())
print([i for i in ids if "aurora" in i])
Two cautions. First, what you see depends on your account, since access to unreleased models is often limited to specific organizations, so an empty result does not mean the reports are wrong. Second, listing an ID is not the same as being able to call it. If you do try a call, the status code tells you something, in general terms:
| Response | Typical meaning |
|---|---|
| 404 not found | The ID does not exist for your account or has been removed |
| 403 forbidden | The ID exists but your account lacks access |
| 429 rate limited | You are sending too many requests |
| 503 service unavailable | The service cannot take the request right now, which can mean a real backend that is out of capacity |
Reports said one aurora ID returned 503 while the other two returned 404. A 503 is weak evidence that something sits behind the name, but it does not tell you what. Do not send sensitive prompts or production traffic to an undocumented model, because you do not know its data handling terms.
A cheap way to keep watch is to run a small scheduled script that records the list of IDs and diffs it against yesterday's. That gives you a timestamp for when any new ID first appears for your account, which is more reliable than relying on social posts.
What each theory would mean if it turns out to be true
If they are image models. Expect an announcement covering pricing per image, resolution and aspect ratio options, text rendering quality, editing and reference-image support, and any safety or provenance features. Teams using an existing image model should prepare a fixed test set of prompts now, so they can compare quality, latency and cost objectively on release day instead of relying on a launch demo.
If they are Codex or coding models. Look for changes to agent harness support, context limits and tool use, plus how they sit against OpenAI's current coding models. A fixed set of your own repository tasks is again the right test.
If they are fast GPT-6 variants. The questions would be speed per token, price per token and where the quality trade-off falls, which our explainer on tokens per second can help you interpret.
In every case, the practical move is the same: write your evaluation first, then wait for documentation.
Why labs test under codenames
Labs often expose candidate models under neutral or colorful names so that testers and users form opinions without the weight of a brand, and so that internal names can change before release. Arena-style testing, where users compare anonymous outputs, depends on this. The downside for outside observers is that a codename can describe a model that is later renamed, merged into another release or dropped, so tracking names alone tells you little about what ships. What eventually reaches documentation is the model that matters.
What we do not know
- Whether the models are image, coding, language or something else
- Whether the IDs are still listed, and to whom
- What red, green and blue stand for
- Pricing, limits and availability
- Whether they relate to the Arena "Aurora" model beyond the shared name
We will update this post if OpenAI documents the models or if the IDs behave differently.
Related reading
- ChatGPT Images 2 and gpt-image-2
- GPT-6 Astra: launch, benchmarks and pricing
- GPT-6.1 Sol: launch, pricing and benchmarks
- How to read AI benchmarks
- ChatGPT Work vs Codex: complete guide
Details come from third-party reports and social posts dated October 17 to 19, 2026, none of which we could verify directly. OpenAI has not confirmed the models.