GPT Image 2

OpenAIimagelive~29stext-to-imageeditingtypography

Text-to-image with dependable in-image typography, plus masked editing from reference images.

from$0.009/ image

Exact price on the Run button. Failed runs cost $0.00.

How this model is billed

Price depends on size and quality: from $0.009 for a low-quality 1024x768 up to $0.369 for a high-quality 3840x2160. The exact figure is shown before you run.

Reserved, then settled to the cent. A failed run costs $0.00.

Typical run ~29s: the measured mean, end to end.

Compare every price

Run GPT Image 2

GPT Image 2 API

Async: the POST answers at once with a request id. Poll it, or take the webhook.

Docs

Submit

Parameters at the top level of the JSON body. Answers 202 with a request id.

curlPOST /v1/models/gpt-image-2

curl -X POST https://api.ezflows.io/v1/models/gpt-image-2 \
  -H "Authorization: Bearer $EZFLOWS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "a wide shot of a desert at dusk",
    "size": "1024x1024",
    "quality": "medium"
  }'

# 202 Accepted, at once — it never waits for the model:
# {"request_id": "req_8f2a41c9b70e4d15", "status": "queued",
#  "poll_url": "https://api.ezflows.io/v1/requests/req_8f2a41c9b70e4d15"}

Poll

Same bearer header, every few seconds, until the status is final.

curlGET /v1/requests/:id

REQUEST_ID=req_8f2a41c9b70e4d15   # from the POST

while :; do
  BODY=$(curl -s https://api.ezflows.io/v1/requests/$REQUEST_ID \
    -H "Authorization: Bearer $EZFLOWS_API_KEY")
  STATUS=$(echo "$BODY" | jq -r .data.status)
  case "$STATUS" in succeeded|failed|cancelled) break ;; esac
  sleep 3
done

echo "$BODY" | jq -r '.data.output.artifacts[].url'   # media.ezflows.io, kept 90 days

Response

Files are on media.ezflows.io, kept 90 days. A failed run says why and costs nothing.

json200 · succeeded

{
  "data": {
    "id": "req_8f2a41c9b70e4d15",
    "model_slug": "gpt-image-2",
    "status": "succeeded",
    "http_status": 200,
    "cost": 0.009,
    "auth_method": "api_key",
    "error": null,
    "submitted_at": "2026-09-12T18:07:21.114Z",
    "queue_started_at": "2026-09-12T18:07:21.402Z",
    "execution_started_at": "2026-09-12T18:07:22.010Z",
    "first_byte_at": "2026-09-12T18:07:52.871Z",
    "finished_at": "2026-09-12T18:07:55.203Z",
    "io_deleted_at": null,
    "canvas_id": null,
    "canvas_node_id": null,
    "client_attempt_id": null,
    "idempotency_key": null,
    "input": {
      "prompt": "a wide shot of a desert at dusk",
      "size": "1024x1024",
      "quality": "medium"
    },
    "output": {
      "artifacts": [
        {
          "index": 0,
          "kind": "image",
          "contentType": "image/png",
          "url": "https://media.ezflows.io/outputs/<account>/req_8f2a41c9b70e4d15/0.png"
        }
      ]
    }
  }
}

Webhook

Skip polling: we POST the result to your endpoint, at least once.

Register: Create an endpoint under Webhooks in the dashboard; the signing secret is shown once.
Events: request.succeeded · request.failed · request.cancelled
Headers: ezflows-event-id (dedupe on it) · ezflows-timestamp (unix seconds) · ezflows-signature t=<ts>,v1=<hex HMAC-SHA256 of "<ts>.<raw body>">
Verify: Over the exact bytes received, in constant time; reject anything older than five minutes.
Respond: Any 2xx within 10 s. Redirects are not followed.
Retries: Five, with backoff: 30 s · 60 s · 2 min · 4 min · 8 min. Disabled after five consecutive dead deliveries.

jsonDelivery body

{
  "id": "82cf46a4-841e-41d6-b3c7-facc937d3252",
  "type": "request.succeeded",
  "created_at": "2026-09-12T18:07:55Z",
  "data": {
    "id": "req_8f2a41c9b70e4d15",
    "model_slug": "gpt-image-2",
    "status": "succeeded",
    "cost": 0.009,
    "output": {
      "artifacts": [
        {
          "url": "https://media.ezflows.io/…/0.png"
        }
      ]
    }
  }
}

curlverify.sh

# Create the endpoint under Webhooks in the dashboard; the signing secret is shown once.
# ezflows-signature: t=<ts>,v1=<hex HMAC-SHA256 of "<ts>.<raw body>">
# body.json = the exact bytes received. Reject anything older than five minutes.
EXPECTED=$({ printf '%s.' "$TS"; cat body.json; } \
  | openssl dgst -sha256 -hmac "$EZFLOWS_WEBHOOK_SECRET" -r | cut -d' ' -f1)

[ "$EXPECTED" = "$V1" ] && echo "valid" || echo "reject"

Parameters 9

Name	Type	Default	Allowed	Description
promptrequired	string	—	1–4000 chars	What to generate. Handles in-image text across scripts.
size	enum	1024x1024	1024x768 · 1024x1024 · 1024x1536 · 1920x1080 · 2560x1440 · 3840x2160	Size
quality	enum	medium	low · medium · high	High costs roughly 3.5x medium at the same size.
image_urls	string[]	—	image urls	Supplying any image switches to edit mode.
mask_image_url	string	—	one image url	White regions are the ones edited. Needs a reference image.
background	enum	opaque	opaque · transparent · auto	Background
output_format	enum	png	png · jpeg · webp	Format
output_compression	integer	100	0–100	Keep at 100 when the image contains text.
moderation	enum	auto	low · auto	Moderation

File inputs (image_urls: image urls · mask_image_url: one image url) take URLs. Upload the raw file to POST /v1/uploads with its own Content-Type and pass the url it returns. Inputs are kept 90 days.

About this model

## Overview GPT Image 2, from OpenAI, generates images from text and edits them from reference images. Its standout is typography: it renders text inside the image dependably, across scripts, which makes it a strong choice for posters, packaging, signs, UI mock-ups and anything else where the words have to be right. Give it a prompt alone and it generates. Add one or more images and it switches to edit mode, and add a mask to confine the change to part of the picture. ## What it does well - **Legible in-image text.** Quote the exact words you want and they are set into the picture. - **Masked editing.** Supply a mask where the white regions are the ones to edit, and the rest of the image is left alone. A mask needs a reference image. - **Transparent backgrounds.** Set the background to transparent for cut-out assets, or leave it opaque. - **Six sizes.** From 1024 by 768 up to 3840 by 2160, including a 1920 by 1080 and a 2560 by 1440 option. - **Three quality tiers.** Low for drafts, medium for most work and high for final art. ## Prompting tips Lead with the subject and the style, then quote any text you need exactly as it should appear. Keep the compression setting at 100 when the image contains text, so the lettering stays crisp. Use PNG when you need transparency or perfectly sharp text, and WebP or JPEG when file size matters. For an edit, describe the result you want rather than listing every pixel that should stay. ## Limits to know Prompts can be up to 4,000 characters. High quality costs roughly three and a half times medium at the same size, and the size you choose moves the price too, so the Run button shows the exact figure for your settings before you spend. ## FAQ ### How do I edit an existing image? Add it as a reference image. As soon as an image is supplied the model switches to edit mode, and an optional mask limits the change to the white regions. ### Can it make a transparent PNG? Yes. Set the background to transparent and keep the output format PNG.