Image Generation Upgraded: gpt-image-2 Is Here
Our AI Image Studio just got a serious upgrade. We have swapped out OpenAI's gpt-image-1 for the brand new gpt-image-2 -- their state-of-the-art image generation model released on April 21, 2026. Within twelve hours of launch it claimed the number one spot across every category on the Image Arena leaderboard by a 242-point margin. That is not a typo.
What Makes gpt-image-2 Different
This is not an incremental update. OpenAI describes gpt-image-2 as a "visual thought partner" rather than a creative toy, and after testing it extensively, that framing holds up. Here is what actually changed.
Text That Actually Works
Two years ago, asking any AI image model to render text inside an image was a guaranteed disaster. Restaurant menus came back with "burrto" and "enchuita." Diagrams had gibberish labels. Those days are over. gpt-image-2 renders accurate, multi-line, multilingual text directly inside generated images. This matters enormously for anyone creating social graphics, flyers, signage mockups, or infographics -- which is exactly what our team uses the Image Studio for.
Flexible Resolution Up to 4K
The old model was locked to a handful of fixed sizes. gpt-image-2 accepts any resolution where both dimensions are multiples of 16 pixels, up to 3840px on the longest edge. We have mapped this into three intuitive tiers in the UI:
- 1K -- fast drafts and social thumbnails (1024x1024, 1792x1024, etc.)
- 2K -- balanced quality for presentations and internal comms (2048x2048, 2048x1152)
- 4K -- maximum detail for print-ready assets and hero imagery (3840x2160, 2160x3840)
The aspect ratio dropdown now combines with the resolution tier, so picking "16:9" at "4K" gives you a proper 3840x2160 landscape image.
Quality Control
A new Quality selector gives you explicit control over the rendering pipeline:
- Low -- fastest generation, cheapest, great for rapid iteration and thumbnails
- Medium -- production-quality social and web assets
- High -- print-ready, maximum fidelity
- Auto -- lets the model decide based on your prompt complexity
At low quality, a 1024x1024 image costs roughly $0.006. At high quality, roughly $0.21. That is genuinely competitive with stock photography, except every image is unique and tailored to your prompt.
Output Format Choice
You can now choose your output format:
- PNG -- lossless, best for graphics with text or sharp edges
- JPEG -- faster generation, smaller files, ideal when speed matters
- WebP -- smallest file size, excellent for web use
JPEG and WebP also support a compression parameter on the backend for fine-tuning file size versus quality.
Reference Image Support
The Reference Images input is now available for all providers, not just Gemini. gpt-image-2 processes every input image at maximum fidelity automatically -- you cannot downgrade the input quality, which means edits are always working from the best possible source material. Upload a source image and describe what you want changed.
What This Means for the Team
The Image Studio at /ai/images works exactly as before -- same layout, same workflow, same Team Feed gallery. The dropdowns now expose more capability without adding complexity. Pick your provider, set your quality and resolution, write your prompt, hit Generate.
The default provider is now gpt-image-2. The previous gpt-image-1 remains available in the dropdown as a legacy option. Gemini 3 Pro Image Preview and Gemini 2.5 Flash are still there for anyone who prefers Google's models or needs Gemini-specific features like Thinking Mode and Grounded Generation.
Cost Comparison
gpt-image-2 pricing is token-based. Larger resolutions and higher quality use more output tokens, but the model is significantly cheaper than its predecessor at every tier:
| Quality | 1024x1024 (1K) | 1024x1536 (1K Portrait) | 2048x2048 (2K) | 3840x2160 (4K) |
|---|---|---|---|---|
| Low | ~$0.006 | ~$0.005 | ~$0.01 | ~$0.03 |
| Medium | ~$0.053 | ~$0.041 | ~$0.10 | ~$0.25 |
| High | ~$0.211 | ~$0.165 | ~$0.40 | ~$1.00 |
You will notice that some non-square resolutions are actually cheaper than square at the same quality -- this is a quirk of how gpt-image-2 calculates output tokens. A landscape 1024x1536 at low quality costs less than a square 1024x1024.
For context, a single stock photo license typically runs $3-10. We can generate dozens of unique, prompt-tailored images at high quality for the price of one stock image.
Technical Notes
- Model default updated in
services/ai-generation.js - All three new parameters (quality, output_format, output_compression) flow through the API route and are logged via TokenWatch under the
ai_generationproject - Resolution mapping handles all valid gpt-image-2 size constraints: multiples of 16px, max 3840px edge, max 3:1 aspect ratio, total pixels between 655,360 and 8,294,400
- The model index has been updated to reflect the change
Happy generating.
-- Reginald