Some ChatGPT users are generating images with a model OpenAI hasn't announced. Reports of the unreleased "Image V2" reaching regular accounts have been circulating for roughly two weeks, since three anonymously codenamed variants surfaced briefly on LM Arena in early April before being pulled.
The 21:9 tell
One circulating test: ask ChatGPT to generate an image at a 21:9 aspect ratio. GPT Image 1.5, the current public model from OpenAI's December launch, doesn't handle that shape reliably. The unreleased one apparently does.
Whether that proves you've been routed to a different model is a separate question. Aspect ratio support could be a routing toggle rather than a signature of anything new underneath. But the trick lines up with what early users flagged when the mystery models first appeared on Arena: sharper composition, clean text rendering, and none of the yellow color cast that dogged the previous version.
What we actually know
On April 4, three models named maskingtape-alpha, gaffertape-alpha, and packingtape-alpha appeared on LM Arena. Whoever picked those names was clearly browsing a hardware aisle. They vanished within hours and returned later with adjustments, and Adam Holter noted that new "Image v2" references had begun surfacing in mobile app code strings. That kind of product integration work usually precedes a broader rollout.
This is OpenAI's playbook. Before GPT Image 1.5 shipped in December, the same Arena blind-testing method used the codenames Chestnut and Hazelnut. Weeks later the model went public.
Who actually has access
Here the reporting starts to fray. An April 19 report claimed a staggered rollout had begun for ChatGPT Plus and Pro subscribers. Russian-language tech channels, meanwhile, are reporting that free accounts are also being pulled in. Both can't be entirely true as stated, which suggests what's actually happening is an A/B test that largely ignores subscription tier.
OpenAI has said nothing. The API documentation still lists gpt-image-1.5 as the current model. Any service claiming you can call "GPT Image 2" programmatically right now is either confused or selling something.
The yellow problem
The color cast complaint keeps coming up. GPT Image 1.5 outputs, and GPT Image 1 before it, had a persistent drift toward warm, vaguely jaundiced tones regardless of prompt. Testers of the tape models kept noting, almost as an aside, that the new outputs looked neutral. Whether that's a new architecture, different training data, or a post-processing tweak is impossible to say without OpenAI disclosing anything. And they won't.
DALL-E 2 and 3 are reportedly scheduled to shut down in mid-May, which creates a natural deadline for OpenAI to publicly clarify what its image lineup actually looks like. That date is worth watching more than any specific launch rumor.
