GPT Image 2.5 API TWD Pricing: flare/sunburst, $30/M Output

Query date: 2026-09-09 (Taipei time). Pricing follows the USD unit prices shown on the BazaarLink models page, and the estimates use 1 USD = NT$ 31.5. This article only covers specifications verified on the site; sizes are currently limited to 1K and 2K.

OpenAI's new-generation image model GPT Image 2.5 was released today, and BazaarLink listed two variants the same day: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both belong to the same generation, have the same price and the same interface, and you can switch model names in existing OpenAI-compatible code to call them.

Bottom line up front: This is the newest OpenAI image model currently available on the site. It supports text and image input with image output, and a 400K-token context. It suits scenarios where you need to edit while looking at reference images (changing the background of product photos, extending a style, multi-turn retouching). Unit prices: input $8 per M, text output $8 per M, image output $30 per M tokens, and cached read $2 per M. Generation cost mostly falls on image output tokens, so focus on that line when estimating.

Want to call GPT Image 2.5 with billing in NT$ and a tax-ID invoice? BazaarLink provides an OpenAI-compatible route. Model prices are the same as the official ones, with a 10% top-up service fee, which addresses the expense-reporting pain point for Taiwan companies. See the complete guide to AI API expense reporting in Taiwan for details.

Specifications at a glance

How do flare and sunburst differ? They are two variants of the same generation, with the same price and the same interface on BazaarLink. If OpenAI publishes an explicit difference in positioning, we will update this article and the model page. Until then, we suggest generating a few images from the same prompts with each and choosing the one that matches your style.

Price comparison (USD ↔ NT$)

Compared with the previous-generation GPT 5.4 Image 2 on the site: image output is the same at $30 per M, input is the same at $8 per M, text output drops from $15 per M to $8 per M, and the context expands from 272K to 400K.

The prices in this table match the BazaarLink models page, with no model markup. For the latest prices, see BazaarLink Models.

How much does one image cost? How to estimate

The cost of an image model is not a "fixed price per image." It is image output tokens × $30 per M. The token count per image varies with size and content, and the actual figure is the usage field in the API response. Estimation method:

• First run 3 to 5 images at 1K size and note the average output token count from the response usage
• Monthly cost ≈ average tokens × images per month × $30 ÷ 1,000,000, then multiply by 31.5 to convert to NT$

Example (assuming a 1K image is about 1,000 output tokens): One image is about $0.03 (NT$ 0.95). At 200 images per day, or 6,000 images per month, that is about $180 (about NT$ 5,670), or about NT$ 6,240 with the 10% top-up service fee. This is only a hypothetical value; replace it with your own usage measurements.

If you are doing multi-turn retouching (feeding the previous image back in each round), input also accumulates. Put the reference image and fixed instructions in the prefix so that cached reads ($2 per M) absorb the repeated portion.

How to call it

Method 1: /v1/images/generations (simplest)

curl https://api.bazaarlink.ai/v1/images/generations \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-flare",
    "prompt": "Hand-drawn poster-style illustration of a Taiwanese night market stall, warm tones",
    "size": "1024x1024",
    "n": 1
  }'

size accepts 1024x1024 (corresponding to 1K) and the 2K level. Put the aspect ratio in aspect_ratio (for example "16:9"), not in size.

Method 2: /v1/chat/completions with a reference image (image-to-image, multi-turn retouching)

curl https://api.bazaarlink.ai/v1/chat/completions \
  -H "Authorization: Bearer $BAZAARLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-sunburst",
    "modalities": ["image"],
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Replace the background of this product photo with a pure white studio backdrop, keep the product details"},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
      ]
    }]
  }'

For Python using the OpenAI SDK, you only need to change base_url to https://api.bazaarlink.ai/v1 and model to either ID above. The rest of the code stays the same.

When to choose it

• Editing while looking at a reference image: changing product backgrounds, extending a style, or producing a finished image from a sketch. Plain text-to-image also works, but this model's strength is image input.
• Multi-turn conversational retouching: the 400K context can keep the full retouching history in the same conversation.
• Budget-sensitive, high-volume image generation: measure usage at 1K first, confirm the cost per image, and then scale up. Image output at $30 per M is the main expense.

Using GPT Image 2.5 in Taiwan: invoices and expense reporting

Top-ups through BazaarLink can be invoiced to a company tax ID as an electronic uniform invoice, with NT$ pricing, so companies do not need to handle overseas receipts for expense reporting. See How to fill in a tax ID for OpenAI? A complete guide to buying OpenAI API invoices in Taiwan for details.

Sources and query log

• BazaarLink model catalog /api/v1/models?output_modalities=image (image models only appear with this parameter; query date 2026-09-09): the two model IDs, input and output modalities, context length, and input and text output unit prices. The image output price of $30 per M and the cached read price of $2 per M are route pricing; the catalog pricing field does not display them separately, so refer to the models page
• BazaarLink API documentation: /v1/images/generations and the modalities parameter, and the size / aspect_ratio rules
• BazaarLink production test (2026-09-09): one generation request was sent to each variant, both produced images successfully, taking 14.4 seconds and 16.3 seconds

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API