GPT Image 2: pricing, image generation and editing

What GPT Image 2 does - and what the name means

GPT Image 2 generates new images from written instructions and edits supplied images. These are different starting points: generation builds a scene from a description, while editing asks the model to change an existing visual. A product image, for example, can become the reference for a different background rather than being described from scratch.

The API model name is gpt-image-2. Searches for “chatgpt image 2” often concern image creation inside a conversational interface, but an API model identifier is not a consumer subscription name. Do not use API token prices to estimate the cost of a ChatGPT subscription, or assume that an image interface exposes a particular model merely because it can generate pictures.

OpenAI’s release history includes later GPT Image 2.5 variants, so GPT Image 2 should not be described as the newest named generation. Its practical relevance is its combination of generation, editing, image inputs, flexible sizes, and Batch processing, not its position in a version-number ranking.

High-fidelity image handling makes reference-led workflows possible. It does not guarantee an exact copy of a face, product label, or material texture. Treat a generated edit as a new image that needs inspection, especially when the unchanged parts carry factual or commercial significance.

GPT Image 2 specifications and API cost calculation

As of September 2026, OpenAI’s pricing tables show two different GPT Image 2 image-token schedules. Keep them separate and match the rate to the endpoint and billing arrangement used by your integration. Neither schedule is a fixed per-image price.

The developer pricing table uses the $4/$1/$15 schedule. For an illustrative request using 1,000 uncached image-input tokens and 2,000 image-output tokens, those image charges total $0.034: $0.004 for input plus $0.030 for output. Under the separate $8/$2/$30 schedule, the same hypothetical usage totals $0.068. These examples exclude text-token charges and other applicable costs.

Cached image-input tokens cost 75% less than uncached image-input tokens under either schedule. That is an input discount, not a 75% discount on the whole request. Output generation still contributes its own charge.

Budget for revisions, not just the first result. Record actual usage for a representative job, then multiply by expected image volume and revision count. Flexible sizing also does not mean arbitrary dimensions: validate your requested size against the selected endpoint before building a production workflow.

A repeatable GPT Image 2 generation and editing workflow

A useful workflow separates the brief, the image request, and the acceptance check. This prevents successive edits from drifting away from the original requirement.

1. Define the deliverable. Specify its use, composition, required text, and intended proportions. A square product card and a wide editorial header need different framing, even when they show the same subject.
2. Choose generation or editing. Use text-to-image for a new concept. Supply an image when an existing object, person, layout, or scene should guide the result. Attach only reference material you have permission to use.
3. Separate changes from protected details. State what should change and what should remain unchanged. For a product photograph, identify the packaging, logo, color, and visible label as preservation requirements.
4. Set supported output controls. Choose the size and quality options available through your endpoint. Save those settings alongside the prompt so that another request can reproduce the same setup.
5. Generate and inspect. Check wording, object counts, anatomy, edges, shadows, and the details carried over from references. Compare edits with the original rather than judging the result in isolation.
6. Revise one issue at a time. Ask for a specific correction, then recheck protected details. For queued jobs that do not require immediate interaction, consider Batch processing and its 50% discount.

A concrete editing prompt might read: “Replace the gray background with a pale blue studio backdrop. Keep the bottle shape, cap, label wording, and camera angle unchanged. Add a soft shadow beneath the bottle. Do not add props.” This defines both the edit and its boundaries without burying the important instructions in decorative adjectives.

Who GPT Image 2 suits, and where manual editing helps

GPT Image 2 suits developers who want generation and editing inside an API-driven workflow. Reference inputs are useful when the task starts with an existing asset; Batch support is relevant when many requests can be queued rather than handled interactively.

• Content teams: create illustration concepts and alternate compositions, then verify that the image does not introduce misleading details.
• Product teams: explore backgrounds and presentation styles while comparing the output against the original product photograph.
• Designers: develop visual directions before rebuilding final typography, spacing, or brand elements in a controlled editor.
• Developers: log prompts, request settings, token usage, and revision counts to understand the cost of accepted results.

It is less suitable as an unchecked final step for images where every detail must be exact. A plausible label is not necessarily the correct label. A convincing diagram can still show the wrong number of components or an impossible connection.

After generation, use conventional editing for deterministic corrections such as cropping or placing approved text. Pict.AI is one option: an AI photo editor app for iPhone and Android, alongside a website with guides and free image tools. Other mobile and desktop editors can serve the same finishing role; this does not imply that they provide GPT Image 2 access.

Alternatives: compare task fit, billing units and licenses

Start with the job you need to complete, not a universal model ranking. Compare candidates using the same brief and acceptance criteria, and count the revisions required to reach a usable result.

Within OpenAI’s family, gpt image 1.5 has image-token rates of $8 input, $2 cached input, and $32 output per million tokens. Against GPT Image 2’s $4/$1/$15 developer-table rates, the input rates are halved and the output rate is about 53% lower. This is a billing comparison, not evidence of a matching improvement in image quality.

If seedream 4.0 is on your shortlist, compare its reference-led workflow with the newer Seedream 4.5 rather than treating their capabilities as interchangeable. Seedream 4.5 supports up to ten reference images, which matters for multi-source compositions.

For text-heavy artwork, put ideogram ai and qwen image on a task-specific shortlist and inspect exact spelling, line breaks, and layout. Qwen-Image-Edit supports Chinese-English text editing with preservation of fonts, sizes, and styles. Those editing features belong to the named editing model, not automatically to every Qwen interface.

A grok image generator workflow may suit conversational generation and editing; Grok Imagine can chain those operations through server-side tooling. Compare reve image when layout-first composition is important, and include google imagen 4 as another candidate without assuming that its pricing or controls match OpenAI’s.

FLUX.2 offers a different cost structure: [klein] 4B starts at $0.014 per image, while [pro] starts at $0.03 per megapixel for generation and $0.045 per megapixel for editing. Token, image, and megapixel prices cannot be compared directly without a representative workload.

Safety, publication rights and final image checks

Technical access to an image model does not settle the rights to every input or output. Before uploading a reference, check whether it contains private information, an identifiable person, a protected brand asset, or material licensed only for a limited purpose.

Before publication, distinguish a visual concept from a factual representation. An edited product image should not invent features or accessories. A generated person should not imply a real endorsement. Images used as evidence, news illustration, or documentation need particular care because a believable result can still be false.

• Rights: review copyright, trademark, likeness, privacy, and contractual restrictions for both references and intended use.
• Accuracy: inspect text, counts, spatial relationships, anatomy, and identity preservation at the final display size.
• Provenance: retain the original assets, prompt, model identifier, settings, and edit history.
• Disclosure: label synthetic or materially altered imagery where the publication context or platform requires it.

Licenses also affect alternative-model selection. FLUX.2 [dev] is designated for non-commercial use, while FLUX.2 [klein] 4B has open weights under Apache 2.0. Neither fact removes the need to assess third-party rights in the image itself.

For GPT Image 2, review the terms applicable to your account and access route before commercial publication. Keep policy compliance and legal clearance separate from visual quality: an attractive image is not, by itself, a cleared asset.