GPT-image-2 leads benchmarks

OpenAI’s GPT-image-2 has moved to the top of OpenAI’s current image stack and into Microsoft Foundry, where Microsoft lists it as the best option for high-resolution and 4K image generation. (learn.microsoft.com; developers.openai.com) Image models turn text prompts — and sometimes existing pictures — into new images or edits. OpenAI’s documentation says GPT-image-2 can both generate images from scratch and edit images through the Image API and the Responses API. (developers.openai.com; developers.openai.com) Microsoft’s Foundry documentation says GPT-image-2 is in public preview and supports text-plus-image inputs, improved editing, broad aspect ratios, and output sizes up to a 3,840-pixel long edge, which Microsoft labels as 4K. (learn.microsoft.com) That changes where the model can be bought and deployed. Microsoft says Foundry offers hosted access to OpenAI models inside its enterprise platform, with pay-as-you-go and provisioned-throughput options rather than a single consumer app entry point. (azure.microsoft.com; learn.microsoft.com) OpenAI’s own API docs now present “GPT Image 2” as the image-generation path for developers, and the platform docs link “Generate images as output” directly to GPT Image 2. (developers.openai.com; developers.openai.com) The model also fits a broader handoff in the market. Microsoft’s Foundry guide says DALL-E 3 was retired for new deployments on March 4, 2026, and tells customers to use the GPT-image series instead. (learn.microsoft.com) OpenAI’s tooling docs describe GPT-image-2 as part of a newer family that can work inside conversational workflows, where a main model can decide whether to generate a fresh image or edit one already in the thread. The same docs say the tool can revise prompts automatically and return base64-encoded image output. (developers.openai.com) Microsoft’s Foundry page also draws a sharper product line around cost and capability: GPT-image-2 is positioned for high-resolution output and editing, while GPT-image-1-mini is described as the cheaper option for fast prototyping and bulk generation. (learn.microsoft.com; developers.openai.com) Arena said on X that GPT-image-2 posted record margins on its image leaderboards, including gains of as much as +242 points in some categories, but the underlying benchmark page was not publicly accessible in my search, so I could not independently verify Arena’s full methodology or category breakdown from a primary source. (x.com) What is verifiable is the distribution shift: OpenAI now documents GPT-image-2 as its latest image model, and Microsoft has already slotted it into Foundry with public-preview access and 4K support. (developers.openai.com; learn.microsoft.com)