llmrelay
// gpt

GPT Image 2

OpenAI’s current image generation model, billed per token.

GPT Image 2 is OpenAI’s current image generation model, and the one to use when DALL·E 3 is not cutting it. It handles both photorealistic and illustration styles, follows complex prompts more reliably than its predecessor, and can render legible text inside images — something earlier models struggled with.

Unlike text models where you estimate token usage from word count, image models bill per output image token, and those tokens are determined by the size, quality, and complexity you request. A 1024×1024 standard-quality image costs about 272 output tokens; larger sizes and hd quality cost proportionally more.

We sell it at half OpenAI’s list price on the text-input tier — $2.50 per million input tokens against $5.00. Image inputs (for edits or variations) cost more at the source ($8/M), but at our rate ($2.50/M) you are paying 31% of list, which is better than 50%. Output tokens are $15/M against OpenAI’s $30, a clean 50% discount.

Official list
llmrelay
Prompt tokens / M
$5.00
$2.50
Image tokens / M
$30.00
$15.00
One image (1024×1024, standard quality)
$0.0082
$0.0041

What it costs you per batch

Priced from a real call through llmrelay: a 1024×1024, standard quality image bills 272 image tokens plus your prompt. Larger sizes and higher quality settings produce more tokens and cost more.

Volume
Official
llmrelay
10 images
$0.08
$0.04save $0.04
100 images
$0.82
$0.41save $0.41
1,000 images
$8.21
$4.10save $4.10

Call it at POST /v1/images/generations with model: "gpt-image-2". Responses come back as base64 in data[0].b64_json.

What it's good at

Best for

Pick GPT Image 2 when

  • +Product visuals, marketing imagery, or social media graphics where you need consistent style and prompt following. GPT Image 2 is demonstrably better at instruction adherence than DALL·E 3.
  • +Images with text — labels, signs, product packaging, infographics. Earlier models mangled text; GPT Image 2 renders it legibly most of the time.
  • +Programmatic image pipelines where you are generating hundreds or thousands of images. At $0.004 per 1024×1024 standard image, volume work becomes feasible.
  • +You need to iterate on a concept. The per-image cost is low enough that generating 10 variations to find the right one is not a budget decision.

Choose something else when

  • !You need fine control over composition, lighting, or style and have the skills to use traditional tools. Photoshop or Blender give you precision that a text prompt cannot match.
  • !The images are load-bearing in a legal or regulated context. Generated images can contain artifacts, biases, or unintended elements that you would not catch until they are published.
  • !You need images of real people in specific contexts. OpenAI’s safety filters block prompts that could generate realistic depictions of identifiable individuals.
  • !Extreme photorealism where any hint of generated-ness is unacceptable. GPT Image 2 is good, but a professional photographer with a real camera is still better for work where authenticity is the whole point.

Questions people ask about GPT Image 2

How much does GPT Image 2 cost per image?

It depends on size and quality. A 1024×1024 standard-quality image costs about 272 output tokens plus a small prompt token cost, which works out to roughly $0.004 per image on llmrelay. Larger sizes (1536×1024, 1024×1536) and hd quality cost proportionally more. OpenAI’s direct pricing would charge about $0.008 for the same image, so you are paying half.

Can GPT Image 2 edit existing images?

Yes, through the edits and variations endpoints. You upload an image (which costs image-tier input tokens at $8/M list, $2.50/M here) and describe what to change. The model generates a new image incorporating your edits.

Is GPT Image 2 better than DALL·E 3?

On instruction following and text rendering, yes — demonstrably. On style range and creativity it is a matter of taste. Run the same prompt through both and pick the output you prefer. DALL·E 3 is still available if GPT Image 2 is not giving you what you want.

Can I use GPT Image 2 for commercial work?

Yes, subject to OpenAI’s usage policies. You own the images you generate. That said, read the policy — there are restrictions around generating images of public figures, violating intellectual property, or creating content that could mislead.

Try GPT Image 2 at half the price

Free to create an account, no subscription, $10 minimum top-up — enough to run this model against a real task and judge quality yourself.

Get API key →

Compare GPT Image 2 against the alternatives

The comparisons this model appears in, the models nearest it on price, and the full rate card.