GPT Image 2.5 Explained: sunburst and flare, xhigh/max Quality, Arbitrary Resolutions

Around 2026-09-08 the OpenAI image family quietly gained a new generation: GPT Image 2.5. There is no browser-readable announcement page to point at — as of 2026-09-09 the OpenAI docs domains (developers.openai.com / platform.openai.com) return 403 to this site — but the evidence sits in the official SDK type definitions: the release notes of both openai-python v3.10.0 and openai-node v7.12.0 carry the same feature line, add GPT Image 2.5 models and image options. Every fact below is checked against those GitHub sources, re-fetchable today (full list in the references at the end); anything that could not be verified — pricing, the positioning split between the two models — is labeled as such rather than guessed.

1. What is GPT Image 2.5?

GPT Image 2.5 is the newest generation of the Images API generation family, with two models:

• gpt-image-2.5-sunburst
• gpt-image-2.5-flare

Each has a dated snapshot: gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08. All four IDs are in the ImageModel enum of openai-python v3.10.0, and they appear in all three model lists: images.generate (text-to-image), images.edit (image editing), and the Responses API image_generation tool.

One boundary matters: SDK types answer "which models exist and which parameters they accept" — nothing more. The positioning split between the two models (which one is the flagship), pricing, and any accompanying announcement are absent from the SDK, and as of 2026-09-09 could not be verified elsewhere because the docs domain returned 403. Until official guidance is reachable, treating the two as a single capability tier is the safer default.

2. The model roster: two generations side by side

The ImageModel enum in openai-python v3.10.0 (source fetched today) now lists:

Two takeaways. First, the gpt-image-2 models were not removed, so existing code is unaffected. Second, seeing dall-e-2 / dall-e-3 in the enum does not mean they are callable — they were hard-removed from the API on 2026-05-12 (see our OpenAI Models Release Notes (2026, Living Document)); the enum entries are type-level residue.

3. The new capabilities, strictly per the SDK

3.1 quality: xhigh and max join the ladder

The SDK describes quality in layers: GPT image models support low / medium / high; gpt-image-2.5-sunburst and gpt-image-2.5-flare (including the 2026-09-08 snapshots) additionally support xhigh and max; the default is auto. The full Python union reads standard (DALL·E legacy) / hd (DALL·E legacy) / low / medium / high / xhigh / max / auto.

The SDK gives no quantified cost or latency curve per tier (experience-based: higher tiers mean more detail, more latency, and more cost). Compare tiers on the same prompt before committing.

3.2 size: three hard rules for arbitrary resolutions

For the 2.5 and gpt-image-2 models, the SDK accepts arbitrary resolution strings WIDTHxHEIGHT (its own example is 1536x864) with three explicit constraints:

1. Width and height must both be divisible by 16;
2. The aspect ratio must stay between 1:3 and 3:1;
3. The ceiling is 3840x2160, and resolutions above 2560x1440 are experimental.

A catch-all applies on top: the requested size must satisfy the model's current pixel and edge limits. The standard sizes 1024x1024 / 1536x1024 / 1024x1536 keep working, and auto remains supported. Valid custom examples: 2048x1152 (16:9, both divisible by 16), 1920x1088 (1088 divided by 16 is 68), 1024x3072 (the 1:3 extreme).

3.3 background: transparency graduates on 2.5

The background parameter takes transparent / opaque / auto (default). The wording difference is worth quoting precisely: for the two 2.5 models (snapshots included) the SDK states outright that they support opaque and transparent backgrounds, while for gpt-image-2 and gpt-image-2-2026-04-21 the same capability is marked in preview. With transparent, the output format must be png or webp.

3.4 Everything else carries over

The images.generate parameter set (source of the day): prompt (required), background, model, moderation (low / auto), n, output_compression, output_format (png / jpeg / webp), partial_images, quality, response_format, size, style (vivid / natural, DALL·E legacy), user, and stream. images.edit adds image (required), mask, and input_fidelity (high / low), and its quality values skip the DALL·E-only hd. None of these were announced as new or removed for 2.5 — the "image options" in the release note are exactly the three items above: quality tiers, size rules, and the background note.

4. Hands-on code (SDK-source-verified)

The samples below are line-checked against the openai-python v3.10.0 parameter types. Honest label: this is a documentation-verified version, not a live-fire one — this site had no usable API key at writing time; every parameter name and value can be pointed to in the sources from section 3.

4.1 Text-to-image: 16:9 high resolution

from openai import OpenAI
import base64
import pathlib

client = OpenAI()

result = client.images.generate(
    model="gpt-image-2.5-sunburst",
    prompt="minimalist tea brand key visual, cream background, one camellia branch, soft side light",
    size="2048x1152",      # both divisible by 16, 16:9, below 2560x1440
    quality="high",
    n=1,
)

pathlib.Path("out.png").write_bytes(base64.b64decode(result.data[0].b64_json))

4.2 Transparent background

result = client.images.generate(
    model="gpt-image-2.5-flare",
    prompt="mascot sticker, single subject, no background",
    background="transparent",   # must pair with png / webp
    output_format="png",
    quality="xhigh",
    size="1024x1024",
)

4.3 Same call from Node

import OpenAI from "openai";

const client = new OpenAI();

const result = await client.images.generate({
  model: "gpt-image-2.5-flare",
  prompt: "isometric SaaS dashboard icon, soft gradients",
  size: "1536x864",
  quality: "high",
});

const b64 = result.data[0].b64_json; // same shape as the Python side

4.4 Image editing

result = client.images.edit(
    model="gpt-image-2.5-sunburst",
    image=open("product.png", "rb"),
    prompt="replace the background with a plain gradient, keep the subject and its shadow",
    input_fidelity="high",
)

5. Migrating from GPT Image 2 to 2.5

Migration path: upgrade the SDK (Python at least v3.10.0, Node at least v7.12.0), swap the model string, keep everything else; check output formats where you rely on transparency; compare old and new outputs on the same prompts before switching over. gpt-image-2 and its snapshot are still in the enum with no deprecation marking as of 2026-09-09 — the arrival of 2.5 is not a reason to rush.

Our GPT Image 2 hands-on (GPT Image 2 API in Practice: From Text-to-Image to Image Editing) covers editing and mask workflows that apply to 2.5 unchanged; prompt engineering method lives in GPT Image 2 Prompt Formulas and Commercial Use Boundaries.

6. What this article does not claim: pricing and positioning

Following our sourcing discipline, the following were unverified on the day and get no assertion here:

• Pricing: the docs domains returned 403 on 2026-09-09 and SDK types carry no prices; do not assume GPT Image 2 pricing carries over.
• The positioning split between the two models: the SDK holds IDs and capability text only — no flagship/lite distinction.
• The official announcement and changelog text: once the docs domains are reachable again, re-check the GA status and accompanying notes; our OpenAI Models Release Notes (2026, Living Document) will get the entry.

How to self-check: from a network that can reach the official docs, open the image pricing page and the images docs on platform.openai.com and confirm whether the model list includes 2.5 and at what price.

7. Common errors and troubleshooting

• Unknown model error: your SDK is too old. Versions before openai-python v3.10.0 and openai-node v7.12.0 do not list the 2.5 IDs — upgrade first.
• The divisibility-by-16 trap: 1920x1080 is invalid (1080 is not divisible by 16); use 1920x1088. 3840x2160 is valid but sits in the experimental range (above 2560x1440).
• Transparency not applying: check that output_format is png or webp — jpeg has no alpha channel.
• No base64 in the response: the GPT image family returns result.data[0].b64_json; do not parse it like the old DALL·E url flow.
• dall-e-2 / dall-e-3 in the enum but calls fail: type-level residue; both were hard-removed on 2026-05-12. Move to the GPT Image family.

8. Next steps

• GPT Image 2 API in Practice: From Text-to-Image to Image Editing — editing, masks, and multi-image composition that apply to 2.5 unchanged.
• GPT Image 2 Prompt Formulas and Commercial Use Boundaries — the five-part prompt formula and a Usage Policy checklist.
• GPT Image 2 vs Midjourney vs DALL·E 3: Three Image Generation Models in Practice (2026) — a selection-oriented comparison.
• openai-python 3.9/3.10 and openai-node 7.11/7.12: Prompt Cache Diagnostics, API Key Expiry, GPT Image 2.5 — the full read on this SDK batch.
• OpenAI Models Release Notes (2026, Living Document) — the timeline view of model releases.