Transparent backgrounds coming to GPT-Image-2 API

OpenAI launched a preview version that adds support for transparent backgrounds in its image-generation model GPT-Image-2 via the API. The feature lets users generate graphic assets that can be placed directly on any background without a separate background-removal step.

Developers who responded to the announcement on X said they have been receiving transparent backgrounds since Image-2’s launch, but consistency is an issue. In fifty percent of calls the model returns a checkerboard pattern—a simulated transparency shown in editing software—instead of a true alpha channel. The asset appears correct in the preview, but in production a gray-white background must be removed manually. The supported output formats include PNG with an alpha channel, which provides genuine transparency.

The launch is labeled as a preview, meaning OpenAI is still collecting feedback and expects to improve reliability before a stable release. In the meantime, anyone building a pipeline that relies on guaranteed transparency must add client-side validation or fall back to conventional background removal. No official benchmarks have been published that detail the exact success rate by object type or prompt.

The direction is correct; native transparency can eliminate a processing step and shorten time to production, but as long as the model produces a checkerboard pattern instead of a real alpha channel in a substantial portion of calls, it remains a tool that requires supervision rather than a plug-and-play solution.