OpenAI’s GPT-image-2 is now generally available on Microsoft Foundry, and it’s a meaningful leap forward for AI-powered image generation.
What’s New?#
- 4K resolution support with an 8.29M pixel budget and custom dimensions
- Multilingual text rendering across Japanese, Korean, Chinese, Hindi, and Bengali
- Intelligent routing that auto-selects the optimal generation config, no manual sizing needed
- Enhanced thinking that lets the model search the web, self-check outputs, and generate multiple images from one prompt
- Knowledge cutoff of December 2025 for more contextually accurate results
This isn’t just an incremental update. GPT-image-2 shifts image generation models from simple tools into genuine creative sidekicks.
Resolution: The Details That Matter#
GPT-image-2 introduces serious flexibility in output dimensions:
| Constraint | Detail |
|---|---|
| Maximum pixels | 8,294,400 (total pixel budget) |
| Minimum pixels | 655,360 |
| Supported resolutions | 4K, 1024x1024, 1536x1024, 1024x1536 |
| Dimension alignment | Each dimension must be a multiple of 16 |
If your requested resolution exceeds the pixel budget, the service automatically resizes it down. No errors, no guesswork.
Intelligent Routing: Two Modes#
Mode 1: Legacy Size Selection maps to three familiar tiers: smimage (small), image (standard), and xlimage (large). Perfect for teams already using the legacy sizes who want automatic selection without code changes.
Mode 2: Token Size Bucket Selection offers six token buckets (16, 24, 36, 48, 64, 96) that map approximately to the legacy tiers but provide finer-grained control. This approach optimizes output quality and efficiency on a per-prompt basis.
See the Difference#
The quality improvement across generations is striking. Using the same base prompt, an empty subway car interior with realistic lighting, the progression from GPT-image-1 through GPT-image-1.5 to GPT-image-2 shows dramatically improved fidelity, detail, and realism.
Even more impressive is the iterative editing capability. Starting from that subway scene, you can add a cohesive ad campaign in the ad frames, then refine those ads to show specific flower types. All in three simple prompts. Concept to production mockup in minutes.
Enterprise Use Cases#
- Retail & E-commerce: Product imagery at exact platform dimensions, no post-processing
- Marketing: Localized campaign visuals at scale
- Media & Entertainment: Storyboard panels at production-pipeline resolutions
- Education & Training: Visual learning aids formatted for any device
- UI/UX Design: Interface assets at precise design-system dimensions
GPT-image-2 ships with Microsoft’s responsible AI safeguards and Azure AI Content Safety built in.
👉 Deploy on Microsoft Foundry | Try the Playground | Read the Docs