Now loading...
Recent indications suggest that OpenAI may be in the process of developing a new image model, tentatively referred to as GPT-Image-3. Over the weekend, three new model references—aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha—were momentarily visible in OpenAI’s API. Shortly thereafter, these models began appearing in blind arena tests focused on image generation.
The discovery was made through an alert on the Dev Mode Discord server, although the references were removed within hours, implying an inadvertent launch.
Analysis from Lumina indicated that all three models are being assessed on Arena, with initial shared outputs including a couple of cat images generated by aurora-green-alpha. While these images displayed a good level of detail, it’s observed that they do not yet exhibit notable artistic quality.
Lumina Bench speculated that the array of models likely belongs to GPT Image 3, a deduction that remains tentative. Testing results revealed that the red model operated at a faster pace but produced lower-quality images, whereas the green model showcased superior aesthetics albeit with slower performance. The blue model has yet to undergo testing.
Additionally, an observer on X named Intius identified three new entries labeled Aspen, Cypress, and Maple on Design Arena shortly after the original model references surfaced, though it’s unclear if they’re directly connected to the Aurora models.
The initiative to experiment with multiple image generation models is consistent with OpenAI’s previous strategies. The company has previously released ChatGPT Images 2.5, which featured two models optimized for different tasks: speed and precision. Earlier in the year, OpenAI also trialed various Image V2 models in a similar fashion.
While the new image generation models are intriguing, many users are eagerly anticipating the launch of a new frontier model. Recently, OpenAI introduced GPT-6 with an Intelligent UI in ChatGPT, following the deployment of the Grok Bot called Dots, which utilizes GPT-6 Astra and GPT-6.1 Sol. Additionally, OpenAI recently added gpt-rosalind-discovery to its pricing page, although it remains unannounced, despite the Rosalind Workbench entering research preview shortly after.
In parallel, competitors such as Anthropic are reportedly preparing to release Claude Fable 5.5, which select partners already have access to. As OpenAI gears up for potentially unveiling GPT-Image-3 or another significant development, the upcoming weeks promise to be pivotal. Currently, however, the details surrounding the Aurora models, including documentation and pricing, remain undisclosed, leaving audiences with limited insights beyond test images.
