GPT Image 2.5: How to use it, Flare vs Sunburst, and API pricing

To use GPT Image 2.5, ask ChatGPT to create an image or call the Images API with gpt-image-2.5-flare or gpt-image-2.5-sunburst. Choose Flare for quick iterations; choose Sunburst when precise edits matter more than waiting time. This guide walks through a product-photo brief, a targeted edit, and the checks to make before exporting.

OpenAI released ChatGPT Images 2.5 on September 8, 2026 (opens in a new tab), with rollout across ChatGPT, ChatGPT Work, and Codex plans. It turns text and reference images into new images and edits. You can also open VizGPT's image workspace (opens in a new tab) to work on your own image briefs.

Name

Soren Atelier

Published on 9/8/2026

Updated on 9/21/2026

Cover: OpenAI's official ChatGPT Images 2.5 example. (opens in a new tab)

What changed in GPT Image 2.5?

OpenAI reports up to 50% lower generation latency than Images 2.0, along with better reference-subject preservation, natural lighting, textures, and consistency across successive edits. That latency figure is a vendor comparison, not a guaranteed time per image. The release also adds Sketch, image comments, templates, and prompt sharing in ChatGPT. Source: OpenAI announcement (opens in a new tab).

The practical question is whether changing one detail leaves everything else usable. A product photo that loses its label after a background edit still needs another revision.

Official example: preserve the portrait, change the clothing

Images: OpenAI's Images 2.5 release example (opens in a new tab), rehosted for this article. These are official examples, not our generated results. Compare the face, blue backdrop, and printed-photo edges as well as the changed clothes.

Flare or Sunburst: which should you use?

This is a workflow recommendation based on OpenAI's positioning of Flare (opens in a new tab) and Sunburst (opens in a new tab). Both accept text and images and support low, medium, high, xhigh, max, and auto quality settings. A higher setting is worth comparing on your hardest image, rather than applying to every draft.

What OpenAI actually lists as different

Checked against the official model pages on September 20, 2026. The two variants share every published parameter; the only stated differences are positioning and the speed and performance labels.

Three practical consequences:

• There is no bare gpt-image-2.5 model ID. The API enum only contains the two variant names and their -2026-09-08 snapshots. A request with model="gpt-image-2.5" fails with a model-not-found error.
• OpenAI's own token estimator has a single entry, "GPT Image 2.5 (Sunburst and Flare)". OpenAI does not publish a per-variant token table. Third-party tests published on September 8 and 18 report identical usage.output_tokens for the same prompt on both variants, but that is a third-party observation, not an OpenAI guarantee. Check the usage field of your own responses.
• Pick by workflow, not by price. With identical rates and, so far, identical token counts, the trade is time against editing precision. Start on Flare; move a brief to Sunburst only when an edit keeps disturbing parts you asked it to preserve.

Sources: Flare model page (opens in a new tab), Sunburst model page (opens in a new tab), image generation guide (opens in a new tab).

Create a product image, then make one edit

1. Write the image brief

Open ChatGPT (opens in a new tab) and ask it to create an image. If you need a specific subject preserved, upload its reference photo. Start with one deliverable and describe the composition before adding decorative detail. OpenAI's image prompting guide (opens in a new tab) recommends explicit text, reference roles, and preservation constraints.

Here is an original prompt you can adapt:

Create a square studio product photograph of a cobalt-blue ceramic mug.
Place the mug slightly right of center, handle facing right.
Use a pale gray background and soft light from the left.
Leave the upper-left quarter empty for a headline I will add later.
Keep the ceramic texture visible. Do not add text, logos, or extra objects.

Adding the headline afterward makes it easier to revise campaign copy without regenerating the photograph. If the image must contain text, quote the exact wording and specify its location.

2. Change only one thing

Continue with the selected image in the same conversation. Try this edit:

Change only the background from pale gray to warm beige.
Preserve the mug's shape, cobalt-blue color, handle direction,
position, ceramic texture, lighting, and shadow.
Keep the upper-left space empty. Add no new objects or text.

If spatial instructions are hard to express, type @Sketch in ChatGPT to draw a layout reference. Image comments can identify the part you want changed. These controls are described in the release announcement (opens in a new tab).

3. Check the result before another revision

Save the version you approve before making the next change. Our checklist is a suggested review procedure; it is not a benchmark result.

Use GPT Image 2.5 with the API

For a single image, use the Images API. For an application that carries image edits across a conversation, use the Responses API image-generation tool. The image model belongs in the tool's model field in that second approach, alongside a supported mainline model at the request's top level. Official API guide (opens in a new tab).

Install the Python SDK in your environment:

python -m pip install --upgrade openai

Set OPENAI_API_KEY in your environment, then save this as generate_image.py:

import base64
from pathlib import Path
from openai import OpenAI
 
client = OpenAI()
result = client.images.generate(
    model="gpt-image-2.5-flare",
    prompt=(
        "Create a square studio product photo of a cobalt-blue ceramic mug. "
        "Pale gray background, soft light from the left, handle on the right. "
        "Leave the upper-left quarter empty. No text or extra objects."
    ),
    size="1024x1024",
    quality="medium",
)
 
Path("mug.png").write_bytes(base64.b64decode(result.data[0].b64_json))

Run the script to request an image and write the returned data to a PNG file:

python generate_image.py

Switch the model string to gpt-image-2.5-sunburst to compare the same brief. This example was checked against the September 8 documentation; our live request stopped at API authentication, so we do not claim a successful generation test. If authentication fails, check your key and API project access before changing the prompt.

What does GPT Image 2.5 cost?

The official model pages list the following rates for both Flare and Sunburst, in US dollars per million tokens:

Sources: Flare pricing (opens in a new tab) and Sunburst pricing (opens in a new tab). These are token rates, not a flat price per image. OpenAI says the GPT Image 2 calculator does not estimate 2.5 token consumption, so do not use an older model's per-image estimate as a 2.5 quote.

How many tokens does one image use?

OpenAI publishes the per-image estimate as an interactive calculator in the image generation guide (opens in a new tab), under "GPT Image 2.5 and GPT Image 2 output tokens", rather than as a table. The values below are read from that calculator on September 20, 2026 for the three recommended sizes. Cost is output tokens × $30 per million; it excludes the text prompt, any reference images, and streamed partial images.

How to read this against the previous generation: in the same calculator, GPT Image 2 at 1024×1024 uses 196 / 1,756 / 7,024 tokens for low / medium / high. So GPT Image 2.5 high costs what GPT Image 2 medium did, and 2.5 max matches the old high. The two new settings, xhigh and max, are where the price climbs; use them on the final campaign image, not on drafts.

Worked example for the mug brief above at quality="medium", 1024×1024: 439 output tokens × $30 / 1,000,000 = about $0.013 per image, plus roughly 60 text input tokens × $5 / 1,000,000 ≈ $0.0003 for the prompt. An edit that sends the previous image back as a reference adds image input tokens at $8 per million; OpenAI has not published the per-reference token count for 2.5, so measure it from usage.

Extra costs and limits to plan for:

• Partial images while streaming: each partial image adds 100 output tokens ($0.003).
• Responses API: the mainline model's own input and output tokens are billed on top of the image tokens.
• Free tier: both model pages list "Free: Not supported"; Tier 1 gets 5 images per minute, Tier 5 gets 250.
• Organization verification: the guide says you may need to complete API Organization Verification before GPT Image models work, so do that before debugging a 403.
• Auto quality: auto picks a level per image, so the estimate above needs an explicit quality value; log usage.output_tokens on every request if you need a real per-image cost.

Start your next image brief in VizGPT

Open VizGPT's image workspace (opens in a new tab), choose the model shown in the generator, and adapt the mug brief to your own subject. Start with a usable composition, select the result you prefer, and make one change at a time.

For data-heavy visuals, keep chart values in a plotting tool and use image generation for surrounding illustration. Our Python visualization guide covers libraries for charts whose geometry must reflect actual data.

FAQ

What are the GPT Image 2.5 API model names?

Use gpt-image-2.5-flare or gpt-image-2.5-sunburst. The model pages also list dated September 8, 2026 snapshots.

Is GPT Image 2.5 always twice as fast?

No. OpenAI reports up to 50% lower latency compared with Images 2.0. That does not guarantee a fixed speedup for every prompt or setting.

Which model should I start with?

Start with Flare for quick drafts. Compare Sunburst when preserving detail through edits is the main requirement.

Is Sunburst more expensive than Flare?

Not per token. Both variants are billed at $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens, and OpenAI's estimator treats them as one entry. Cost differences come from the quality setting, size, and edit inputs you choose.

How much does one 1024×1024 image cost?

Using OpenAI's estimator: about $0.006 at low, $0.013 at medium, $0.053 at high, $0.094 at xhigh, and $0.211 at max, excluding prompt and reference-image tokens.

Related Guides

• Stable Diffusion prompt guide
• Prompt generators for Stable Diffusion
• Python data visualization libraries
• How to use Codex