OpenAI's leaked "Aurora" models hit Arena, hinting at GPT-Image-3

Weekend API updates are where frontier labs usually drop their breadcrumbs, and OpenAI was just caught red-handed again.

Three new model references—aurora-red-alpha, aurora-green-alpha, and aurora-blue-alpha—briefly surfaced in OpenAI's API before getting yanked within hours. The references were spotted via an alert on the Dev Mode Discord server, looking very much like an accidental rollout. But by then it was too late: all three were already popping up in blind arena tests generating images.

Red for speed, green for pixels

Community testers on Arena quickly started putting the models through their paces. Benchmark account Lumina highlighted that the trio are dedicated image models and speculated they likely represent GPT-Image-3.

Early results show a familiar split in responsibilities. The red variant runs fast at the expense of finer fidelity, while green takes its time to generate noticeably sharper, cleaner frames. The blue variant has not been publicly tested yet.

Lumina shared a couple of cat portraits generated by aurora-green-alpha. While they look clean and crisp, a pair of pet pictures hardly constitutes a benchmark, and so far the output does not look wildly revolutionary. About an hour after the API references appeared, X user Intius also spotted three new entries on Design Arena named Aspen, Cypress, and Maple, though nothing officially ties them to the Aurora models yet.

This tiering strategy is straight out of OpenAI's playbook. When the company launched ChatGPT Images 2.5, it split the API offerings into Flare for raw speed and Sunburst for detail—models that later powered virtual try-on tools in ChatGPT shopping. Going back further, OpenAI ran three tape-themed Image V2 variants through LM Arena in the exact same fashion in April.

The model rush heats up

While community members are having fun rolling prompts on Arena, the broader expectation remains fixed on raw frontier intelligence. OpenAI recently shipped GPT-6 with Intelligent UI in ChatGPT, following up on Dots, its Grok Bot and Muse competitor running on GPT-6 Astra and GPT-6.1 Sol. Just last week, gpt-rosalind-discovery briefly appeared on OpenAI's pricing page right before Rosalind Workbench launched into research preview.

Meanwhile, Anthropic is knocking at the door with Claude Fable 5.5, which enterprise partners reportedly already have and Cursor's newest builds have begun referencing. OpenAI clearly needs to keep shipping to maintain momentum.

Why it matters

OpenAI has yet to release documentation or pricing for the Aurora models, meaning the community is piecing clues together from arena cat images. Still, this testing cycle tells us a lot about where the ecosystem is heading.

For creators and developers, the speed versus quality tradeoff is finally solidifying into formal API tiers. Running real-time design agents or e-commerce prototyping demands rapid, cheap passes like red, while production assets can afford the latency of green. If Aurora indeed becomes GPT-Image-3, users will likely get a unified architecture that scales smoothly between instant ideation and polished final delivery.

For the wider industry, the battleground has shifted from raw aesthetic appeal to latency engineering and cost structures. OpenAI knows it cannot win enterprise pipelines on quality alone if generating a high-res image stalls an interactive workflow. Testing varied compute budgets on Arena allows them to calibrate consumer patience against operational costs.

That said, nothing we have seen from the initial outputs suggests a monumental architectural leap. If GPT-Image-3 ends up being just an incremental tweak wrapped in clever tiered packaging, it may struggle to excite a market that is eagerly waiting for the next genuine breakthrough.