GPT Image 1.5: image generation, editing and API costs

What GPT Image 1.5 does - and what its name means

GPT Image 1.5 is an image model, not a conventional photo editor with a fixed set of sliders. You describe the result you want, supply an existing picture when the task calls for one, and receive a generated image. The API identifier uses hyphens: gpt-image-1.5. That identifier matters when selecting a model in an application or configuring an API request.

OpenAI describes improvements in instruction following and prompt adherence. Those are relevant to requests containing several conditions: change the background, retain the subject, preserve a particular color, and avoid adding text. Better adherence does not mean every condition will be satisfied, especially when instructions conflict or the source image contains small, ambiguous details.

The model supports image generation and editing through dedicated API endpoints. Editing tasks can include adding or removing elements, combining visual content, and changing a scene while attempting to retain important reference details. A generated edit still needs inspection; facial likeness, product geometry and lettering should not be treated as automatically preserved.

GPT Image 1.5 is not the newest named generation in OpenAI’s image family. For a fresh integration, compare its cost and behavior with newer options rather than choosing it solely because an existing tutorial uses that model ID. For an established workflow, evaluate any replacement on the actual images and instructions you need to process.

How text prompts and reference images shape an edit

A text-only request describes a new image. An editing request also provides visual material that the model can use as a reference. The practical difference is important: describing a generic blue bottle is not the same task as preserving the exact silhouette, cap and label of a supplied bottle.

Separate your instructions into what must change and what must remain. For example: “Replace the gray backdrop with a pale cream studio background. Keep the bottle shape, cap color, label wording and camera angle unchanged.” This gives the model a clearer target than “make this look professional,” which leaves the intended changes undefined.

For a newly generated scene, specify the subject, composition, setting and visual treatment. A useful prompt might request a ceramic mug on a wooden table, photographed from the side, with soft window light and empty space on the right for a headline. Describe visible requirements rather than stacking vague quality adjectives.

Reference images help communicate appearance, but they do not turn generation into a pixel-preserving operation. Inspect the whole result, including areas you did not ask to change. When revising, use a narrow instruction such as “remove the extra handle” instead of restating a long prompt that could encourage unrelated changes. Save acceptable intermediate images so an unsuccessful revision does not replace your best version.

A six-step workflow for generating and editing images

The most reliable workflow makes the acceptance criteria explicit before generation. Decide whether success means a convincing concept, a faithful product edit or a publishable asset; those are different standards.

1. Choose generation or editing. Start with text for a new scene. Supply an image when the existing subject, layout or visual details matter.
2. Prepare the reference. Use a clear image in which the relevant subject is visible. Remove unrelated or sensitive material that you do not need to submit.
3. Write the change and preservation rules. Identify the desired background, lighting or composition, then list the details that must stay unchanged. Put exact requested wording in quotation marks.
4. Select the model explicitly. In an API workflow, use gpt-image-1.5 rather than assuming a platform’s generic “image” option selects it. Choose the output settings available through your endpoint.
5. Inspect against a checklist. Count objects, read every word, inspect hands and faces, and compare the edited subject with the reference. Check the crop and any space reserved for design elements.
6. Revise narrowly and keep versions. Correct one failure at a time where practical. Retain the prompt, source image and accepted output so you can understand what changed between attempts.

For publication, make a separate final review after the image looks visually acceptable. Confirm that a product has not acquired invented features, a person’s appearance has not been materially altered, and any generated text is correct. A pleasing thumbnail can conceal errors that become obvious at full size.

GPT Image 1.5 pricing and alternatives compared

As of September 2026, GPT Image 1.5 costs $8 per million input image tokens, $2 per million cached input image tokens and $32 per million output image tokens. These are token rates, not a flat price for each finished picture. The OpenAI API pricing page separates image-token charges from other applicable pricing categories.

For arithmetic only, 10,000 output image tokens at $32 per million cost $0.32. That example is not an estimate of one image’s token consumption and excludes input charges. Cached-input pricing is a separate rate; do not apply it to every uploaded reference automatically.

When comparing gpt image 2, check the applicable endpoint’s pricing: OpenAI also lists a $8/$2/$30 image-token schedule on a separate pricing table. Lower token rates alone do not establish a lower cost per accepted image. For flux 2, resolution and variant affect the bill. Evaluate retries, reference handling and the number of usable outputs, not just the headline rate.

Limits to check before using an output

Instruction-following improvements do not remove the usual image-generation failure modes. Text, object counts, spatial relationships, anatomy and identity preservation all deserve explicit review. A request for three objects may produce an extra one; a label may appear convincing while containing a misspelling. Treat these as acceptance checks rather than cosmetic details.

Preservation is especially important in editing. Compare logos, seams, jewelry, facial features and product proportions against the original. If accuracy is contractual or factual, use generation for the surrounding scene and keep critical artwork under direct editing control. For lengthy copy or legally required wording, adding typeset text after generation is generally easier to verify than accepting lettering embedded in the image.

Do not assume that a third-party interface exposes every API capability or uses the same billing structure. Output settings, upload restrictions and access conditions need to be checked for the specific endpoint or application. An app’s credits or subscription are not interchangeable with OpenAI’s image-token rates.

Publication also requires a rights and privacy review. Consider copyright, trademarks, identifiable people, consent and the risk of misleading viewers. A model’s ability to generate something does not establish permission to publish it. Commercial terms differ across alternatives: FLUX.2 [dev], for example, is designated for non-commercial use. Keep licensing decisions separate from judgments about image quality.

Useful projects and when to compare another tool

GPT Image 1.5 is useful for iterative visual concepts: exploring a product against different backgrounds, drafting an editorial illustration, changing a scene’s mood, or developing a campaign image before final design work. These tasks benefit from natural-language instructions and tolerate a review-and-revision cycle. Exact packaging reproduction or tightly controlled brand artwork requires more careful checking.

For multi-source compositions, seedream ai is worth comparing through Seedream 4.5, which supports up to 10 reference images and emphasizes preserving reference details. The useful question is whether it retains your particular subject, not whether a sample gallery looks attractive.

If contextual edits are central to the job, include flux kontext pro in the shortlist. For text-heavy artwork, compare ideogram 3.0 using the exact words and layout you need rather than assuming any model will reproduce a finished design correctly. These comparisons should use the same brief and the same acceptance criteria.

For conversational workflows, grok imagine supports generation and editing calls that can be chained within a conversation. A qwen image editor comparison is relevant when changing embedded Chinese or English text while trying to retain fonts, sizes and styling. That is a more specific task than general image creation.

For straightforward mobile photo edits, Pict.AI is another option: it is a free AI photo editor app for iPhone and Android, with a website offering guides and free image tools. Choose between an editor and a model API based on whether you need a ready-made interface or a programmable workflow, not on model naming alone.