OpenAI quietly lists three Aurora alpha models in its API without announcement or documentation

Three model identifiers called aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha appeared in OpenAI’s API model list on October 17, 2026, without any official announcement, documentation, pricing or changelog from the company.

The discovery was made by API observers and confirmed by authenticated calls to the /v1/models endpoint. According to OpenAI’s own records, the three were created between 01:13 and 01:15 UTC on the 17th, roughly 26 hours before the first public mentions on X.

What appeared and what did not

The “owner” field of the models reads “system”, the same value used for official OpenAI models. None has a deactivation date. Their creation within a window of less than two minutes suggests a single deployment rather than three independent experiments.

As of this writing, OpenAI’s API changelog and official models page do not mention any of the Aurora identifiers. There are no public prices, rate limits or capability descriptions.

Arena tests and speculation

Shortly after appearing in the API, the same names began showing up in blind tests on Arena, where they generated images. Reports from LuminaBench describe aurora-red-alpha and aurora-green-alpha producing cat portraits, with red faster and of apparently lower quality, and green slower with a more polished visual result. The blue model has not been widely tested publicly.

This has led to speculation that they are variants of a forthcoming image-generation model (possibly related to a GPT-Image-3). Other circulating hypotheses on X include high-speed specialized coding models similar to the earlier Codex-Spark, or internal versions of more recent GPT models. None of these interpretations has been confirmed by OpenAI.

Prior context

The name “Aurora” had already appeared in February 2026, when OpenRouter listed a cloaked model called Aurora Alpha, described as a reasoning model optimized for speed. Three days later, OpenAI announced GPT-5.3-Codex-Spark, an ultra-fast coding model. The temporal coincidence led many to associate the two, but neither OpenAI nor OpenRouter confirmed the link.

The first relevant public mention of the new identifiers came from researcher Lisan al Gaib (@scaling01) on X on October 18 at 02:59 UTC, followed by other observers.

What remains unconfirmed

So far, OpenAI has not issued an official statement. The models may be removed, renamed or launched under other names at any time. There is no public information about access, costs or benchmark performance. Any practical use should be treated as experimental and subject to change without notice.

The silent appearance of alpha models in the API is a known pattern in the sector, often preceding announcements or broader testing. In this case, the link to image generation observed in Arena adds a concrete data point to the speculation, but the full picture still depends on official communication from OpenAI.

Sources

Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and available sources on the internet. Consult the original sources to verify the full context.

By GeekikiBot