GPT image 2.5 Sunburst

GPT-Image 2.5 Sunburst: Model Introduction & Practical Guide

GPT-Image 2.5 Sunburst is OpenAI's precision-first image generation and editing model - the quality lane of the GPT Image 2.5 family that shipped on September 8, 2026 alongside its speed-oriented sibling, GPT-Image 2.5 Flare.

Here's the short version: OpenAI split GPT Image 2.5 into two API models instead of one. Flare handles high-volume, latency-sensitive work; Sunburst is built for jobs where a wrong detail costs money - surgical edits, reference-subject fidelity, and multi-turn sessions that must not drift. Both share the same token pricing, the same size envelope, and the same quality dial (auto, low, medium, high, xhigh, max), so the real decision is not price per token but cost per accepted image. Sunburst is the slower model by design: third-party paired runs put it roughly 10-15 seconds behind Flare per image at the high setting, and OpenAI positions it above GPT Image 2 on image quality, with official examples covering identity-preserving outfit swaps, single-object removal, and clean transparent product cutouts.

This guide covers model overview, core features, technical specifications, capability comparison, core advantages, recommended use cases, example prompts, selection recommendations, and pricing/API notes.

---

Quick Facts

Table of Contents

1. Model Overview
2. Core Features
3. Technical Specifications
4. Capability Comparison
5. Core Advantages
6. Recommended Use Cases
7. Example Prompts
8. Selection Recommendations
9. Pricing & API Notes
10. FAQ
11. Sources & Further Reading

---

1. Model Overview

GPT Image 2.5 is not one model with two nameplates - it is a deliberate segmentation of the image API into two products with different jobs. OpenAI introduced the pair on September 8, 2026, and the consumer-facing surface shipped as ChatGPT Images 2.5.

Sunburst takes the precision side of that split. Its improvements over GPT Image 2 concentrate on what happens after the first generation: keeping a person, pet, or product recognizable when it is placed into a new scene or style; changing only the elements the prompt actually asks to change while preserving surrounding composition and brand treatment; and staying stable across multi-turn edits, where earlier models tended to accumulate drift and quality loss.

To be precise about naming: there is no bare gpt-image-2.5 model ID in the API. Integrations must specify either gpt-image-2.5-sunburst or gpt-image-2.5-flare, or pin a dated snapshot such as gpt-image-2.5-sunburst-2026-09-08. That matters for reproducible evaluations - record the exact ID you tested, because "GPT Image 2.5" is a family name, not a callable model.

2. Core Features

Fine-grained instruction following. Sunburst's core promise: change what the prompt names, preserve what it does not. Official comparisons show an outfit swap that keeps the subject's face, expression, and pose, an object removal that does not repaint the scene, and a product extraction that keeps bottle geometry and label detail.

Reference-subject fidelity. People, pets, and products stay recognizable when moved into new scenes, styles, and compositions - the capability behind identity-preserving edits.

Multi-turn edit stability. Repeated modifications build on confirmed results instead of degrading them, which is what makes long editing sessions practical in an API workflow.

Text and structured layout generation. Ads, UI mockups, slide-style visuals, and infographics with readable typography - with the family reminder that fact labels still need human verification.

Transparent asset output. PNG/WebP alpha output for isolated product shots, logos, and cutouts without re-drawing the subject.

Six quality tiers. low, medium, high, xhigh, max, plus auto as the default - a wider quality ladder than GPT Image 2's three tiers.

Custom sizes up to 4K. Output up to 3,840 px on the long edge with aspect ratios up to 3:1; sizes beyond 2560×1440 remain experimental.

Provenance built in. Outputs carry C2PA metadata and an invisible watermark, per OpenAI's disclosure.

3. Technical Specifications

Note: the quality parameter and the model choice are independent. Raising Flare to max still calls Flare - it does not switch you to Sunburst.

4. Capability Comparison

Reading the table honestly: the two 2.5 models differ in behavior, not in pricing or size envelope. Third-party paired runs at the high setting (single runs each, same prompts and references) reported Flare/Sunburst wall times of roughly 20.5/30.7 s, 22.7/38.8 s, and 19.7/29.4 s - directional evidence that Sunburst costs time, not a service-level guarantee.

5. Core Advantages

1. Edits that respect boundaries. The model is tuned to change the named element and leave the rest of the frame intact - the single most valuable property for commercial retouching.
2. Identity survives the move. Subjects remain recognizable across scenes, styles, and composition changes.
3. Multi-turn sessions don't degrade. Confirmed decisions survive subsequent instructions, which shortens review cycles.
4. Priced like the fast model. Sunburst and Flare share token rates, so precision does not carry a premium rate - only a latency cost.
5. One ladder, six tiers. The expanded quality dial (through xhigh and max) lets teams dial cost against detail per deliverable.
6. Built for pipelines. Images API for single jobs, Responses API for conversational editing, with C2PA provenance attached.

6. Recommended Use Cases

• Brand and advertising finals: campaign visuals where headline, layout, and product treatment must survive every revision.
• E-commerce product imagery: transparent cutouts, scene swaps, and label-accurate product shots.
• Identity-preserving edits: wardrobe, scene, or style changes that must keep the subject recognizable.
• Complex retouching: multi-step edit chains (remove, replace, relight) without cumulative drift.
• Structured visuals: infographics, UI mockups, and slide-style graphics where typography and hierarchy matter.
• Comics and storyboards: consistent characters across sequential frames.

7. Example Prompts

1. Identity-preserving wardrobe edit

2. Surgical object removal

3. Transparent product asset

4. Text-accurate campaign image

Prompting guidance: state the change and the invariants in the same prompt. Sunburst responds to explicit preservation language ("keep the face, pose, and background unchanged") far better than to a bare edit instruction.

8. Selection Recommendations

Choose GPT-Image 2.5 Sunburst if:

• An edit must not disturb approved elements - client work, product shots, regulated brand assets.
• Subject identity has to survive scene or style changes.
• You run multi-turn editing sessions where drift is the main failure mode.
• The deliverable includes transparent cutouts or text-accurate layouts.

Choose GPT-Image 2.5 Flare if:

• Throughput and latency dominate - social assets, prototypes, batch exploration.
• Quality comparable to GPT Image 2 is sufficient for the task.

Consider alternatives if:

• You need per-pixel hard masking guarantees - no generative editor provides that; composite the model output in a normal image pipeline instead.
• You need academic-grade micro-typography or the deepest multi-reference fusion; that remains a specialist arena.

9. Pricing & API Notes

Sunburst and Flare share the same published token rates, which is unusual in a two-tier lineup - check the live pricing page before budgeting, but at review time the rates were:

Three practical implications:

• Equal rates do not mean equal per-image cost. The two models can consume different token counts for the same request, and references plus retries add input cost. Compare cost per accepted image, not rate cards.
• Pick the model first, then the quality tier. Switching quality tiers never switches models; a low Sunburst is still Sunburst.
• Integration surface matters. Use the Images API for one-shot generation or editing, and the Responses API (with Sunburst set in the image_generation tool) when you need conversational, multi-turn edits. Verify you receive a decodable image payload, not just a successful request.

Sources & Further Reading

• GPT-Image 2.5 Sunburst model page - OpenAI Platform
• Image generation guide - OpenAI Platform
• OpenAI official site
• GPT Image 2.5 guide: Flare vs Sunburst, specs and editing - iMini
• GPT Image 2.5 Flare vs Sunburst: speed, edit precision and API cost - LaoZhang AI Blog
• GPT Image 2.5 dual-model breakdown: Flare vs Sunburst - StartAI

---

GPT-Image 2.5 Sunburst is a recent release; specifications here reflect OpenAI's published documentation and independent write-ups as of late September 2026. Token pricing, rate limits, and per-image costs vary by account and workload - verify against the live model page before production deployment.