GPT Image 2.5 API: Flare vs Sunburst, 4K Sizes and Pricing (2026)

Short answer

GPT Image 2.5 comes as two API models: gpt-image-2.5-flare for fast, everyday creation and gpt-image-2.5-sunburst for detailed work that needs precise edits. Both use the OpenAI Images API format (/v1/images/generations and /v1/images/edits). For 4K output, pass an explicit size such as 3840x2160: the API rejects 4k as a size string, any side longer than 3840 pixels (so no 4096x4096), and anything above 8,294,400 pixels in total. On TeamoRouter each image is billed by the resolution you actually get back, in 1K, 2K and 4K tiers.

Flare vs Sunburst

OpenAI released GPT Image 2.5 on September 8, 2026. Launch coverage (for example Grida's model page) describes the split like this:

Both share GPT Image 2's token rates: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. Each offers quality settings from low to max, plus auto. A practical way to choose: generate drafts and variations with Flare, then switch the model string to Sunburst for the final image or for edits where details must stay put.

Calling the API

Generate an image:

curl https://api.teamorouter.com/v1/images/generations \
  -H "Authorization: Bearer $TEAMO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-flare",
    "prompt": "A product photo of a ceramic mug on a wooden table, soft morning light",
    "size": "2048x2048"
  }'

With the OpenAI Python SDK, point base_url at the gateway:

from openai import OpenAI

client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="sk-teamo-...")
result = client.images.generate(
    model="gpt-image-2.5-sunburst",
    prompt="A product photo of a ceramic mug on a wooden table, soft morning light",
    size="3840x2160",
)

Edits go to /v1/images/edits and accept multipart/form-data uploads. All parameters are in the docs for gpt-image-2.5-flare and gpt-image-2.5-sunburst.

4K sizes: what works and what returns 400

"4K" here means UHD-class output, not 4096 pixels. In our tests in September 2026, the upstream API accepted only literal WIDTHxHEIGHT sizes with the longest side at most 3840 and at most 8,294,400 pixels in total (3840 × 2160, or 2880 × 2880):

If you're moving code over from GPT Image 2, where 4096 × 4096 worked, change the size first; that's the most common cause of a 400 on 2.5.

How billing works

Images are billed per image, by the longest side of the image you get back, not by what you asked for:

With size set to auto or left out, the model usually returns 1024 or 1254 pixels, billed as 1K or 2K. 1536x1024 counts as 2K because its longest side is above 1024. Pass an explicit size if you need a fixed tier. Current per-image prices for each tier are on the pricing page; they're discounted from the official list rate and change with live discounts.

Paying for it

Image generation uses the same balance as Claude, GPT and the other models on TeamoRouter. Top up by card or Alipay. You can also pay with USDT, USDC or ETH from any exchange app: click Pay with crypto in the top-up dialog, and the checkout shows which networks your amount supports (BEP20 works from $5). No wallet extension needed. See how to pay with crypto.

FAQ

Is 4k a valid size? No. Use WIDTHxHEIGHT, for example 3840x2160.

Why does 4096x4096 work on GPT Image 2 but not on 2.5? GPT Image 2.5 caps the longest side at 3840 pixels and the total at 8,294,400 pixels.

Which model should I use for edits? Sunburst is positioned for detailed work that needs precise edits; Flare for fast, everyday generation.

Can Codex generate images through TeamoRouter? Yes. If image generation fails inside Codex, see fixing Codex image generation on an API gateway.

Related

• GPT Image 2.5 Flare API docs
• GPT Image 2.5 Sunburst API docs
• GPT Image 2 API docs