GPT-Image-2.5 API Guide: Flare vs Sunburst, Pricing & Implementation

01

Overview

OpenAI launched two API models with GPT Image 2.5: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both generate and edit images from text and image inputs, but OpenAI positions them for different production priorities. Flare is the faster everyday choice; Sunburst is the higher-precision option for workflows where tighter editing control matters.

02

Flare vs Sunburst

Use Flare when speed, iteration volume and everyday image generation are the priority. OpenAI describes it as its fastest high-quality everyday image model and specifically points to creator content, social content, product experiences, visual search, rapid prototyping and high-volume generation.

Use Sunburst when edit precision is more important than latency. OpenAI positions it for premium visual workflows such as production-ready campaign creative and polished product imagery.

The models should not be reduced to 'cheap vs expensive': OpenAI currently lists the same token rates for both. The practical tradeoff is speed and editing precision.

03

Official pricing

Current OpenAI model cards list these standard rates for both models:

• Text input: $5.00 per 1 million tokens
• Cached text input: $1.25 per 1 million tokens
• Image input: $8.00 per 1 million tokens
• Cached image input: $2.00 per 1 million tokens
• Image output: $30.00 per 1 million tokens

Text output is not billed because the models return images. OpenAI does not provide one universal official flat cost per finished image for GPT Image 2.5. Resolution, quality, prompt/reference inputs and retries affect total token usage.

04

Model IDs and snapshots

OpenAI documents the undated aliases gpt-image-2.5-flare and gpt-image-2.5-sunburst. It also lists dated September 8, 2026 snapshots. Pinning a dated snapshot is useful when a production workflow needs more stable behavior; using the undated alias is useful when you want OpenAI’s current default version.

05

Quality settings

Both model cards list low, medium, high, xhigh, max and auto quality settings. Higher quality can increase generation work and cost. Do not assume that the same named quality level produces the same token consumption as an older image model; OpenAI explicitly notes that its GPT Image 2 calculator does not estimate GPT Image 2.5 token consumption.

06

Endpoints

OpenAI documents direct image generation through v1/images/generations. Sunburst also documents image editing through v1/images/edits, and the model cards say the 2.5 models can be selected in the Image API or used through the Responses API image-generation tool. Check the live model page before deployment because endpoint support can evolve.

07

A practical model-routing strategy

A production system can start with Flare for drafts, variations, social assets and quick edits. Route an asset to Sunburst when the job has high value and repeated edit fidelity matters—for example, a product shot, campaign key visual or final brand asset. This routing approach follows OpenAI’s positioning but is an implementation recommendation, not an official requirement.

08

Cost controls

Track cost per accepted asset rather than only cost per request. A workflow with lower-quality drafts, caching, fewer unnecessary reference images and fewer failed regenerations can be more efficient than repeatedly generating at the highest setting.

Log the selected model, quality, input references, output size, retries and whether the asset was accepted. That gives you real internal cost evidence rather than relying on generic internet estimates.

09

Editing and reference images

Both 2.5 models accept image inputs. The launch emphasizes preserving subjects and changing only requested elements. For marketing automation, this is useful when a workflow needs to keep product identity, layout or brand treatment stable while changing one part of an asset.

Reference-image handling also means teams should consider privacy and rights. Only send images you are authorized to process, and verify current OpenAI data-handling terms for your account and API agreement.

10

Rate limits

OpenAI says rate limits depend on the organization’s usage tier. Do not publish one universal request-per-minute number as if it applies to every account. Developers should check their organization’s live limits.

11

When to choose each model

Choose Flare for rapid ideation, creator/social output, large variation sets, prototypes and latency-sensitive product experiences. Choose Sunburst for final campaign work, polished product imagery and iterative edits where changing the wrong part of the image has a high cost.

Run your own evaluation set before committing. Include the actual assets your workflow cares about: faces, products, text-heavy layouts, transparent backgrounds and repeated edits.

12

FAQ

Are Flare and Sunburst priced differently?

Not on the current OpenAI model cards. Their listed token rates are the same.

Is there an official fixed per-image price?

OpenAI currently documents token pricing for the 2.5 models rather than one universal fixed price per image.

Can I edit existing images?

Yes. The 2.5 launch and model pages describe image editing and image inputs. Check the live endpoint documentation for the exact integration path you use.

Which model should I start with?

For most high-volume or everyday workflows, OpenAI positions Flare as the default. Use Sunburst when editing precision is more important than speed.

Can I pin a model version?

Yes. OpenAI lists dated September 8, 2026 snapshots alongside the undated model IDs.

13

Sources

Reverify the live OpenAI model cards and pricing before publication because model availability and pricing can change.