GPT Image 2 can generate and edit images for a Photoshop workflow through the OpenAI API. The practical route is either a manual generate-download-place loop or a bring-your-own-key Photoshop panel that sends the request and returns the result as a layer. OpenAI bills API use separately from ChatGPT, and some organizations may need a one-time verification step.
This guide uses OpenAI documentation reviewed July 11, 2026. It separates provider capabilities from Creator MAX interface choices and from Bielfor’s own workflow recommendations.
What is GPT Image 2?
OpenAI describes gpt-image-2 as its current image generation and editing model. The API accepts text and image inputs, supports generation and edit endpoints, and offers low, medium, high, and auto quality settings.
The image generation guide allows flexible dimensions when all of these constraints are met:
- the longest edge is no more than 3,840 pixels;
- both edges are multiples of 16 pixels;
- the long-to-short-edge ratio is no wider than 3:1;
- total output is between 655,360 and 8,294,400 pixels.
OpenAI treats outputs above 3,686,400 total pixels, the area of 2,560×1,440, as experimental. A 3,840 by 2,160 request fits the documented limits, but that does not make every larger request equally predictable.
How much does GPT Image 2 cost?
GPT Image 2 is token-metered. The official pricing page lists the following Standard and Batch rates as reviewed July 11, 2026.
| Meter | Standard | Cached standard | Batch |
|---|---|---|---|
| Text input | $5.00 / 1M tokens | $1.25 / 1M tokens | $2.50 / 1M tokens |
| Image input | $8.00 / 1M tokens | $2.00 / 1M tokens | $4.00 / 1M tokens |
| Image output | $30.00 / 1M tokens | Not applicable | $15.00 / 1M tokens |
The complete formula is: total request cost = text input + image input + image output. Text-to-image requests pay prompt input plus output. Edit requests also pay image-input tokens. Reference images add input cost because GPT Image 2 processes each image input at high fidelity.
The next table recomputes output-only charges from the OpenAI calculator. Each charge equals output tokens multiplied by $30 per million. Prompt and reference-image inputs are additional.
| Requested size | Low tokens | Low output charge | Medium tokens | Medium output charge | High tokens | High output charge |
|---|---|---|---|---|---|---|
| 1024×1024 | 196 | $0.00588 | 1,756 | $0.05268 | 7,024 | $0.21072 |
| 1024×1536 | 158 | $0.00474 | 1,372 | $0.04116 | 5,488 | $0.16464 |
| 2048×2048 | 397 | $0.01191 | 3,568 | $0.10704 | 14,272 | $0.42816 |
| 2048×1152 | 157 | $0.00471 | 1,413 | $0.04239 | 5,650 | $0.16950 |
| 3840×2160 | 371 | $0.01113 | 3,336 | $0.10008 | 13,342 | $0.40026 |
Quality changes cost sharply at a fixed size, but size and aspect ratio matter too. For example, high-quality 2,048-square output is $0.42816, while high-quality 1,024-square output is $0.21072. A non-square request can also use fewer tokens than a square request with fewer total pixels, so use the calculator for the requested dimensions.
May OpenAI require organization verification?
Yes, conditionally. OpenAI’s image generation guide says an organization may need API Organization Verification before it can use GPT Image models. This is not a universal setup requirement for every organization.
If the developer console requests verification, complete it for the organization behind the key. OpenAI’s verification help article says status can take up to 30 minutes to update. It also recommends checking the selected organization and creating a new key if a verified account still reports an error.
The useful troubleshooting order is: read the returned error, confirm billing, confirm the organization selected in the console, complete verification only if requested, wait for propagation, and then retry with a newly created key if needed.
How many input images can GPT Image 2 edit?
OpenAI’s edit endpoint accepts up to 16 images for GPT Image models. The API reference says each supported PNG, WebP, or JPG input must be under 50 MB.
Creator MAX exposes 15 add-on reference slots for GPT Image 2 as a product limit. The number 15 leaves room for the selected Photoshop image, so 15 add-on references plus that selection stay inside OpenAI’s 16-image request cap. Provider documentation establishes 16 total inputs; it does not establish a separate 15-reference rule.
That distinction matters when comparing tools. A different client can allocate the 16 inputs differently. Do not treat Creator MAX’s interface cap as an OpenAI model capability.
How does a layered Photoshop edit work?
A panel can export the selected Photoshop region, send it as an edit input with the prompt and optional references, and place the returned pixels in a new layer. The layer placement can match the selected bounds, but model behavior inside those bounds still needs review.
OpenAI explains in the mask section of its image guide that GPT Image uses a mask as prompt guidance and may not follow the mask shape with complete precision. The same guide documents remaining limits in precise text placement, recurring-brand consistency, and layout-sensitive composition.
For label work, retain protected source pixels when possible, inspect every character, and compare the result with the approved packshot before delivery. A correctly positioned layer is not proof that the generated content is correct.
Which GPT Image 2 quality should a Photoshop user choose?
Bielfor workflow judgment, not a provider claim: start with medium for a representative production test, use low for disposable composition drafts, and move to high only when the visible quality gain justifies the measured cost. Compare at the final delivery size, not only at 100 percent zoom.
Bielfor builds Creator MAX, so this recommendation comes from the product owner. The plugin uses the user’s OpenAI key and reserves one of the 16 image inputs for the active Photoshop selection. Other scripts and panels can implement a different workflow.
What happens to API prompts and images?
OpenAI says API data is not used to train its models unless a customer opts in. Its data-controls guide also says default abuse-monitoring logs can retain customer content for up to 30 days. The image generation and edit endpoints are eligible for approved Zero Data Retention controls, subject to OpenAI’s requirements.
A local Photoshop history does not mean the provider receives no data. Prompts and image inputs must leave the device to reach the OpenAI API. Review provider retention, client permissions, and internal handling rules before submitting confidential material.
FAQ
Do I need ChatGPT Plus for GPT Image 2 API use?
No subscription is required for API access. OpenAI’s ChatGPT Plus article lists Plus at $20 per month and states that API usage is separate and billed independently.
Is GPT Image 2 priced per image?
Not as one flat fee. OpenAI meters text input, image input, and image output tokens. The output-only table above is useful for comparison, but an edit total also includes prompt and image-input charges.
Does GPT Image 2 support 4K-class output?
The documented size list includes 3,840 by 2,160 and 2,160 by 3,840. Both fit the 3,840-pixel edge and 8,294,400-pixel total limits. Outputs above 3,686,400 total pixels, the area of 2,560×1,440, are currently experimental.
Is organization verification mandatory for every OpenAI account?
No. OpenAI says an organization may need verification. Follow the console prompt and returned API error for the organization connected to the key.
Are 15 references an OpenAI limit?
No. OpenAI documents up to 16 image inputs on the edit endpoint. Creator MAX uses one for the active selection and exposes 15 add-on references as its current interface limit.
