OpenAI may be testing three variants of a future image model
Hidden references to aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha briefly circulated in OpenAI's API before appearing in blind tests. No official announcement yet, but the signal points to a new image model.
OpenAI appears to be preparing a new image model, likely named GPT-Image-3, following the brief appearance of three API references — aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha — this past weekend. According to observations shared by TestingCatalog, these entries were then removed, resembling an accidental deployment. Subsequently, they appeared in blind tests on Arena, a detail that matters as it suggests advanced trials, even if nothing is confirmed at this stage.
Trois noms, une piste encore floue
The three references spotted in OpenAI's API were reported by a Dev Mode Discord community alert, then removed shortly after. TestingCatalog notes that the coincidence with image tests on Arena did not escape observers, while Lumina published generation previews, including a second cat portrait produced by aurora-green-alpha. The results appear clean and detailed but are insufficient to judge the model's actual level. For now, the clues point to an image system under testing, not a black-and-white announced release.
Splitting into three variants is not unusual at OpenAI. The company has already offered two image models in ChatGPT Images 2.5, with GPT-Image-2.5 Flare for speed and GPT-Image-2.5 Sunburst for accuracy. The tests observed here thus resemble a known logic: confronting several versions of the same system to arbitrate between speed and rendering. Early feedback shared online suggests aurora-red-alpha is faster but renders less well, while aurora-green-alpha is slower and considered higher quality. Aurora-blue-alpha has not yet been publicly tested according to these sources.
Pourquoi ce signal intéresse le marché
This type of leak mainly says something about OpenAI's pace. The company multiplies internal references and field tests while keeping a tight schedule for announcements. Meanwhile, it already listed gpt-rosalind-discovery on its pricing page last week without an official announcement yet, then opened Rosalind Workbench in preliminary research the next day. The signal is clear: several projects are advancing in parallel, and image models remain a strategic area, notably for shopping, design, and creative generation uses.
OpenAI is not alone in this race. Anthropic is reportedly preparing Claude Fable 5.5, according to partners cited in sources, and Cursor has already started referencing the name in its latest version. In this context, the coming weeks could mainly serve as a test of nerves among players refining their models before exposing them to the public. For users, this means one very simple thing: new features often first appear as traces in an API, not as grand announcements. Rumors travel faster than press releases, which has almost become a sport.
Key points
- Three references: aurora-red-alpha, aurora-green-alpha, aurora-blue-alpha.
- OpenAI removed them shortly after their appearance in the API.
- TestingCatalog spotted the alert on the Dev Mode Discord.
- Lumina saw the models on Arena on October 18, 2026.
- OpenAI has already tested GPT-Image-2.5 Flare and Sunburst.
- No indication yet of official documentation or pricing.
By the numbers
- 3 models — references spotted in OpenAI's API, according to TestingCatalog.
- 2 variants — GPT-Image-2.5 Flare and Sunburst, already used by OpenAI.
- 1 day — gap between the pricing page for gpt-rosalind-discovery and Rosalind Workbench.
- October 18, 2026 — date of reports cited in sources.