GPT Image 2.6 API (Flare / Sunburst)

Autenticação

Authorization: Bearer YOUR_API_KEY

Pricing: per token

Billing follows the way the model itself is metered — by the tokens the request actually used, with text and image priced separately and separate rates for input, cached input and output. Each request total is rounded up to $0.0001. An upper bound is held from your balance when the request is accepted; when the image is ready you are charged for the tokens used and the rest is released — credits in the task result is exactly what that request cost. Failed requests are not charged.

Flare and Sunburst share these prices. Quality, size, GIF frames, reference / mask / markup images and prompt length all change the token count, so there is no fixed price per image. Without a quality parameter apimodels.app renders at medium (the model's own default is high, which uses more image output tokens).

For reference: measured requests

5 November 2026, 16:9 text-to-image, one image — what each request actually cost, not a price per image.

Option 1: OpenAI SDK (sync)

An OpenAI-style WxH size makes the call synchronous: the response carries the image (b64_json + url), matching images.generate() / images.edit(). Both dimensions must be multiples of 16 (1024x1024, 1536x1024 and 1024x1536 all work).

# Text-to-image, OpenAI Images API shape — a WxH size makes the call synchronous
curl -X POST https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-flare",
    "prompt": "a ceramic coffee cup on a linen tablecloth, morning window light",
    "size": "1536x1024",
    "quality": "medium"
  }'
# → { "created": ..., "data": [{ "b64_json": "...", "url": "https://r2.apimodels.app/..." }] }

# Edit — multipart, OpenAI images.edit() shape
curl -X POST https://api.apimodels.app/v1/images/edits \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "model=gpt-image-2.6-sunburst" \
  -F "image=@product.png" \
  -F "prompt=change only the label to green, keep everything else exactly the same" \
  -F "size=1024x1024"

Option 2: native async (every 2.6 parameter)

Without a WxH size the request follows the platform async contract: create returns a taskId, poll until completed and read resultUrls (one URL per image when you ask for several), or pass callback_url for a webhook. GIF, markup images, masks and num_images all go through this path.

# Animated GIF: 12 frames, 120 ms each
curl -X POST https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-flare",
    "prompt": "a paper boat bobbing on small waves, loopable",
    "aspect_ratio": "1:1",
    "resolution": "1K",
    "output_format": "gif",
    "frame_count": 12,
    "frame_duration_ms": 120
  }'
# → { "code": 200, "data": { "taskId": "..." } }

# Edit guided by a markup image (circles / arrows / notes on a transparent PNG)
curl -X POST https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-sunburst",
    "prompt": "apply the changes marked on the image, keep everything else",
    "image_urls": ["https://example.com/room.png"],
    "markup_image_url": "https://example.com/room-markup.png",
    "reasoning_effort": "max"
  }'

curl "https://api.apimodels.app/v1/images/generations?task_id=TASK_ID" -H "Authorization: Bearer YOUR_API_KEY"

Parâmetros

Sizes

Exact sizes must have both sides in multiples of 16, a long edge of at most 3840, a ratio of at most 3:1 and 655,360 to 8,294,400 pixels in total. Anything else returns a 400 with the reason — the size is never changed silently. aspect_ratio + resolution examples:

With no size at all, text-to-image uses the model default landscape_4_3 and edits follow the input image (auto). aspect_ratio auto or size auto lets the model pick.

Notes and limits

• GIF: all frames are laid out on one sheet whose longest edge can be at most 3840 px (the frame count rounds up to a power of two and fills a grid — 6 frames use 4×2 or 2×4, 9 frames 4×4; a size fits if either layout does, so for 5–8 frames keep the short side at 960 or less and the long side at 1920 or less). With aspect_ratio + resolution (or no size) apimodels.app shrinks the frame size so the sheet fits — 6 frames at 1:1 1K come back at 960x960; with an exact size it is up to you, and 960x960 fits up to 16 frames. Measured on 5 November 2026: 6 frames at 832x832 low cost $0.034 per request, 6 frames at 976x672 low (fits only as 2×4) $0.0486, 9 frames at 960x960 low $0.092.
• SVG is temporarily unavailable: the model's documentation lists svg output, but every SVG request currently fails upstream, so apimodels.app returns a 400 for output_format: svg until it works.
• Markup image or mask, not both. On the previous generation (2.5) our tests showed masks acting as hints rather than hard boundaries; 2.6 has not been measured yet — for a strict local edit, name the region in the prompt as well, e.g. "change only the bottle label, keep everything else unchanged".
• Measured latency (5 November 2026, 16:9 text-to-image): 1K at reasoning low — about 6s on Flare, 8s on Sunburst; at the default high reasoning Sunburst took 11–14s at 1K and 14–16s at 2K. High quality, 4K, GIF and xhigh / max reasoning take longer; for sync calls set client timeouts to at least 240s.
• Not every request succeeds: moderation refusals and upstream errors happen, and failures are not charged. Result files are deleted after 30 days.
• Upload size: apimodels.app accepts about 4.5MB per request body; for large or many references use https://api.apimodels.app/v1 (same key, 100MB per request).
• If you need a fixed, predictable per-image price, use GPT Image 2.5 (fixed price by resolution × quality).

Try it in the Flare / Sunburst playground