GPT-Image-2.6-All Generation & Editing

​

Overview

• Price: $0.025 per image, billed per call; 2K/4K costs the same
• Speed: about 10–46 s (1K 10–22 s, 2K 12–24 s, 4K 15–42 s; tail latency can exceed 60 s)
• Quality: long edge 512–1536 is native — the most controllable ChatGPT web quality and fingerprint; 2K/4K is super-resolution, which solves large-canvas size, not necessarily more detail than 1.5K
• Formats: PNG / JPEG / GIF / SVG, transparent background and 24-hour URLs; GIF / SVG are native to this route and not available on 2.5-all
• Chinese: native Chinese prompts; strong text rendering for posters, signage and infographics

​

Quickstart

​

Text to image (curl)

curl https://api.apiyi.com/v1/images/generations \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-all",
    "prompt": "A beige paper-textured autumn reading club poster, vertical Chinese title in the center, hand-drawn maple leaf at the bottom right",
    "size": "1536x1024",
    "response_format": "b64_json"
  }'

​

Image editing (Python, multipart)

import base64
import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APIYI_API_KEY"],
    base_url="https://api.apiyi.com/v1",
    timeout=600,
)

with open("input.png", "rb") as f:
    result = client.images.edit(
        model="gpt-image-2.6-all",
        image=f,
        prompt="Replace the background with light gray, keep the subject and text",
        size="1536x1024",
    )

image_bytes = base64.b64decode(result.data[0].b64_json)
with open("edited.png", "wb") as f:
    f.write(image_bytes)

​

Capabilities

​

Sizes

​

Native range (≤1536, ChatGPT web native)

• No 16/64 alignment required. 1537, 1600x900 and 1920x1080 are rejected on the native route.
• size=auto is decided upstream. We have seen 1312x1199, 1402x1122 and 1448x1086; no fixed value is promised.
• On the native route, the response size equals the actual output size.

​

Super-resolution range (2K / 4K, large-canvas compatibility)

​

Choosing a size

• Best ChatGPT-native feel / stable quality: keep size at or below 1536; around 1.5K is ideal. This is the most controllable tier, with the ChatGPT fingerprint.
• 2K large-canvas compatibility: use 2048x2048, 2048x1152, 2048x1536; common sizes are exact, but do not expect more detail than 1.5K.
• 4K landscape: use 3840x2160; it usually returns 3840x2160, occasionally 4096x2731 — verify actual pixels.
• Large squares: the cap is 2880; 4096x4096 comes back as 2880x2880.
• 4:1 banners: not supported; 4096x1024 returns 400.

​

Formats

• Transparent background: background=transparent + PNG. At 2K the result is RGBA with about 40% fully transparent pixels. 4K transparent output does not follow size; no dimensions are promised.
• URL: response_format=url works in the Default group; no Image2_OSS group is needed. The link is a 24-hour signed URL.
• No WebP: the error is output_format 'webp' is not supported; use png, jpeg, gif or svg.

​

GIF

curl https://api.apiyi.com/v1/images/generations \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-all",
    "prompt": "A looping animation of a yellow doll frowning, puffing its cheeks and stomping, pure white background, seamless loop",
    "size": "2048x2048",
    "output_format": "gif",
    "frame_count": 8,
    "frame_duration_ms": 150
  }'

• output_format=gif is mandatory. If you only ask for a loop in the prompt and omit the format, the compatibility layer expects PNG and the request fails with 502 image 1 is gif, png was asked for.
• frame_count / frame_duration_ms are honored exactly: 8 frames / 150 ms, 4 frames / 300 ms and 6 frames / 120 ms all matched the request.
• size does not determine the output size: 1K results have included 384x512, 507x506 and 619x623; 2K returned 327x585; 4K returned 621x621. Only frame count and duration are promised.
• To get a fixed GIF size, resize or crop on the client after generation.

​

SVG

curl https://api.apiyi.com/v1/images/generations \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.6-all",
    "prompt": "A flat rocket icon, exactly three colors, no text",
    "size": "2048x2048",
    "output_format": "svg"
  }'

• True vector: the response is <svg> text, not a bitmap. 2048², 3840x2160, 4096², 6144² and 8192² all passed.
• viewBox follows size: 2048x2048 gives viewBox="0 0 2048 2048"; 3840x2160 gives viewBox="0 0 3840 2160".
• No svg object needed: output_format=svg is enough. Official routes or some SDKs add "svg": {"model": "..."}; 2.6-all accepts but ignores that field — it does not fail and does not change the result, so do not depend on it.
• The output contains no <script> or embedded raster image.

​

Image editing

• Endpoint: POST /v1/images/edits, multipart/form-data.
• Single image / mask / multiple references are accepted. A mask is treated as a reference; edits are not guaranteed to stay inside the mask, so describe what to change and what to keep.
• An explicit size is exact, including 2K/4K (2048x2048 and 3840x2160 both verified).
• Without size or with auto, the output is normalized to about 1.57 MP while preserving the input aspect ratio; the original pixel size is not guaranteed.
• Two reference images work in practice. For more, compose them on the client or edit in steps.

​

n and responses

• n supports 1–10 and is billed per image. At 2K, n=2 returned two 2048x2048 images; n=3 returned three images with usage.image_count=3 and output_image_tokens=8160 (3 × 2720). At $0.025 per image that is $0.075; the final charge is whatever the bill shows.
• The default is b64_json; response_format=url returns only url, never both.
• Common response fields: created, data[], usage. The size field may be missing for 2K/4K; 1K usually echoes it.
• URLs are signed for 24 hours; regenerate after expiry.

​

Timeouts and concurrency

• Set the client timeout to 600 seconds: 1K takes about 10–22 s, 2K 12–24 s, 4K 15–42 s, and tail latency can exceed 60 s. Loosen reverse proxies, serverless limits and queue consumers as well.
• Concurrency: at 1K, 12×2, 16×1 and 8×1536 all passed without 429s. 2K/4K has not been load-tested; start with 4–8 concurrent requests and watch tail latency.
• Billing is synchronous: the request is billed when sent; a client disconnect is not refunded and the result is not retained.

​

Limits

• 2K/4K is super-resolution, for large-canvas compatibility. It is not native pixels and is not promised to carry more detail than the 1536 native tier; for stable quality, stay at or below 1536.
• 2K/4K depends on a fallback route. Requests first hit the native route and are switched to the super-resolution route when the size is out of range. If the fallback is unavailable, the API may return 400 wide_fallback_required; retry or stay at or below 1536.
• GIF size is not guaranteed. 1K, 2K and 4K can all return small images; only frame count and duration are promised.
• WebP is not supported. Only PNG / JPEG / GIF / SVG.
• 4:1 is not supported. 4096x1024 returns 400; the maximum aspect ratio is 3:1.
• 4K transparent output is unstable. We have seen several 4096-long-edge heights; do not promise 3840x2160. Keep transparent output at or below 2K.
• Do not rely on quality. low / medium / high are accepted, but low may return 2048x1153 at 2K while medium / high return 2048x2048; no same-seed quality difference has been verified.
• Editing with auto rescales. The output is about 1.57 MP, not the input size.
• Content policy: upstream moderation returns 400 content_policy_violation; revise the prompt and retry.

​

Errors

​

Let an AI agent wire it up

​

FAQ

​

Is 2K/4K native resolution?

​

Why is size missing from a 2K/4K response?

​

Can it produce a 4K transparent image?

​

Can I set the GIF dimensions?

​

Is WebP supported?

​

Does quality matter?

​

Why does 4096x1024 fail?

​

Related

• Text to image API
• Image editing API
• GPT-Image-2.5-All overview
• Official vs reverse comparison
• Image API best practices