OpenAI's gpt-image-2 reasons about the layout before it draws. Multilingual text, high-fidelity edits and faithful instruction following, online and without watermark. 4 credits per image on Aura AI, with GPT Image 1.5 from 2 credits.
GPT Image 2, known in the API as gpt-image-2, is OpenAI's image generation and editing model, released on April 21, 2026 together with ChatGPT Images 2.0. It is the third model in the line after GPT Image 1 (April 2025) and GPT Image 1.5 (December 16, 2025), which made edits more precise and up to four times faster. GPT Image 2 is OpenAI's quality-first follow-up and the first OpenAI image model with a native reasoning step: it plans the composition, the elements and the text before it renders a pixel, which is why it behaves less like a paint tool and more like a designer that reads the whole brief.
The practical gains are in three areas. Text rendering now covers dense paragraphs, small lettering and non-Latin scripts including Chinese, Japanese, Korean, Hindi and Bengali, the weak spot of every earlier OpenAI image model. Resolution moves from three fixed sizes to custom dimensions, with native output up to 3840 x 2160 upstream. Editing changes philosophy too: where GPT Image 1.5 exposed a low or high input fidelity switch, GPT Image 2 processes every input at high fidelity by default, so faces, logos and fine detail survive a change. OpenAI also corrected the warm color cast of 1.5 toward neutral reproduction. In the API, the high quality tier at 1024 x 1024 costs $0.211 per image upstream, against $0.133 for GPT Image 1.5.
On Aura AI, GPT Image 2 runs in the text-to-image tool at a flat 4 credits per image, in five aspect ratios, and as GPT Image 2 Edit in the image editor at 4 credits per edit. GPT Image 1.5 is there too at 2 credits for both modes, and GPT Image 1 at 3 credits. Output is delivered without watermark, and the same credits work on Nano Banana, FLUX 2, Seedream 5.0 and the rest of the catalog.
Six capabilities that separate gpt-image-2 from GPT Image 1.5 and GPT Image 1
GPT Image 2 thinks through the request first: what goes where, which elements matter, how the text fits. Multi-part prompts with several subjects and constraints land far more often on the first try.
Paragraphs, small lettering and mixed-language layouts come out readable, including Chinese, Japanese, Korean, Hindi and Bengali. Menus, packaging and UI mockups no longer need a second pass in a design tool.
Every edit input is processed at high fidelity, with no toggle to remember. A person's face, a product label or a logo stays exactly as it was in the source while the rest of the scene changes.
Instead of three fixed sizes, GPT Image 2 accepts custom dimensions upstream with native output up to 3840 x 2160. Wide banners, tall stories and print layouts are generated at the right shape.
Because it plans structure, GPT Image 2 is unusually good at app screens, dashboards, slides and web mockups that look like real software rather than an impression of it.
OpenAI corrected the warm cast that GPT Image 1.5 was known for. Skin tones, whites and product colors reproduce neutrally, and photorealistic scenes gain in material detail and lighting.
Four steps from prompt to a finished image, or from a photo to a faithful edit
Flat cost per image, no resolution or aspect ratio multiplier
| Variant | Mode | Credits per image | Best for |
|---|---|---|---|
| GPT Image 2 | Text to image | 4 credits | Final assets, dense text, UI mockups, complex briefs |
| GPT Image 2 Edit | Image edit | 4 credits | High-fidelity edits of photos, products and logos |
| GPT Image 1.5 | Text to image | 2 credits | Quick drafts and volume generation |
| GPT Image 1.5 Edit | Image edit | 2 credits | Fast, precise edits at the lowest cost |
| GPT Image 1 | Text to image | 3 credits | The original 2025 model, for matching earlier work |
How OpenAI's image model compares to the other flagship generators on Aura AI
| Feature | GPT Image 2 | Nano Banana | FLUX 2 | Seedream 5.0 |
|---|---|---|---|---|
| Maker | OpenAI | Google DeepMind | Black Forest Labs | ByteDance |
| Latest release | Apr 2026 | Feb 2026 (Nano Banana 2) | Dec 2025 (FLUX 2 Max) | Jul 2026 (5.0 Pro) |
| Credits, text to image | 4 (1.5: 2) | 3 | 1 to 3 | 3 to 4 |
| Credits, image edit | 4 (1.5: 2) | 2 to 3 | 2 to 3 | 2 to 4 |
| Reasoning step | ✓ Native | ✓ Yes | Not advertised | ✓ Since 5.0 |
| Max resolution (upstream) | 3840 x 2160 | 4K | 4 megapixels | 4K |
| Non-Latin text | ✓ CJK, Hindi, Bengali | ✓ Multilingual | Not specified | ✓ Chinese, English |
| Plan | All plans | All plans | All plans | All plans |
Where reasoning, text and faithful edits pay off
Describe a screen and GPT Image 2 lays it out like a product designer: navigation, cards, buttons and readable labels in the right places.
Run one campaign in several scripts. The same poster can carry a Japanese, Korean or Hindi headline with correct characters, which earlier OpenAI models could not do reliably.
Upload a packshot and ask for a new setting or season. High-fidelity input handling keeps the label, typography and brand marks intact while the environment changes around them.
Charts, step diagrams and comparison slides benefit from the reasoning step: hierarchy is planned first, so labels line up with what they describe.
Common questions about GPT Image 2 on Aura AI
OpenAI's reasoning image model, 4 credits per image, high-fidelity edits, no watermark. Available on every plan, worldwide.