GPT Image 2.5 Migration Notes: Flare for Throughput, Sunburst for Precision

I’d benchmark the edit loop before swapping models

The interesting question with GPT Image 2.5 is not whether a first render looks better. It is whether I can change a label, preserve a face, adjust a layout, and still finish with an acceptable asset without repeatedly starting over.

OpenAI released the generation on September 8, 2026, with two API models:

• Flare targets fast, high-quality everyday generation.
• Sunburst targets maximum generation quality and editing precision.

According to OpenAI’s announcement, Flare delivers higher image quality than Image 2 at 50% lower latency. Sunburst spends more time on quality and control. Both retain Image 2’s published token rates.

That makes Flare my starting candidate for a general-purpose migration, not Sunburst by default. But I would treat the latency improvement as a vendor claim to validate against my workload, and the early benchmark lead as a reason to test—not a replacement for testing.

What actually changes in the API choice

“GPT Image 2.5” is a family name. The model IDs expose two different latency-quality trade-offs.

Detail	Flare	Sunburst	Image 2
API model ID	gpt-image-2.5-flare	gpt-image-2.5-sunburst	gpt-image-2
Release	September 8, 2026	September 8, 2026	April 21, 2026
Inputs	Text and images	Text and images	Text and images
Tasks	Generation and editing	Generation and editing	Generation and editing
Quality controls	low, medium, high, xhigh, max, auto	low, medium, high, xhigh, max, auto	Earlier quality controls
Intended trade-off	Throughput and fast iteration	Maximum quality and edit precision	Established production baseline

Flare’s documentation positions it for creator content, social assets, product experiences, visual search, prototyping, and high-volume generation.

Sunburst’s documentation describes it as OpenAI’s most capable image-generation and editing model. I would evaluate it for final campaign assets, product photography, and revisions where changing an unrequested detail is expensive.

Image 2 is still a useful control. Released alongside ChatGPT Images 2.0 in April 2026, it supports text and image inputs, image output, flexible image sizes, and high-fidelity image inputs. Its dated snapshot is:

gpt-image-2-2026-04-21

That snapshot matters if an integration needs pinned behavior rather than a moving alias. The Image 2 documentation remains the reference for existing deployments.

The benchmark lead is real data, but still preliminary

The Arena snapshot used for this comparison places Sunburst first, Flare second, and Image 2 third in overall generation and single-image editing.

These are human-preference signals from the text-to-image leaderboard and image-edit leaderboard, not guarantees about a particular application.

Benchmark	Flare	Sunburst	Image 2
Text-to-Image Overall	1399 ± 13, Preliminary	1421 ± 13, Preliminary	1381 ± 4
Single Image Edit	1491 ± 9, Preliminary	1520 ± 9, Preliminary	1461 ± 3
Text Rendering	1434 ± 22, Preliminary	1481 ± 23, Preliminary	1427 ± 7
Cartoon, Anime & Fantasy	1428 ± 21, Preliminary	1440 ± 21, Preliminary	1400 ± 6
Portraits	1456 ± 35, Preliminary	1450 ± 32, Preliminary	1427 ± 8

The 2.5 entries had materially fewer votes than Image 2 when this snapshot was prepared. I would not treat the ordering or gaps as settled.

Editing is the result I care about most

Sunburst’s single-image-edit score is 1520, versus Image 2’s 1461: a 59-point lead. Flare scores 1491, a 30-point lead.

For a production pipeline, that could matter more than an attractive initial render. An image can be nearly right and still become unusable after a small revision changes the packaging, face, lighting, or surrounding text.

The benchmark does not directly measure my regeneration budget. It does make editing a sensible place to concentrate evaluation effort.

Overall generation and typography favor Sunburst

In overall generation, Sunburst’s 1421 is 40 points above Image 2’s 1381. Flare’s 1399 is 18 points ahead.

For text rendering, Sunburst reaches 1481, compared with Image 2’s 1427, a 54-point difference. Flare scores 1434.

That is relevant to posters, labels, signs, interface elements, and infographics—not just photography-like outputs. I would still test the actual copy and layout constraints my application needs.

Portraits do not establish a clear Flare win

Flare narrowly leads Sunburst in this portrait snapshot: 1456 versus 1450. But the intervals are wide—±35 and ±32—and the vote counts are relatively small.

I would not turn that into “Flare is the portrait model.” It is a reminder that the quality-first option will not necessarily lead every category.

Where I would look for production gains

OpenAI describes more natural lighting, richer textures, better reference preservation, stronger instruction following, and improved handling of complex layouts. Those claims are most useful when translated into concrete failure cases.

Local edits without collateral changes

The release announcement specifically highlights targeted edits that preserve surrounding details.

My test case would be something like replacing a bottle label while keeping the bottle shape, table, camera angle, lighting, and background intact. If the model redraws the entire scene, the requested edit may succeed while the workflow fails.

Sunburst is the obvious candidate when that precision is the main acceptance criterion. Flare still deserves comparison, given its preliminary editing score and lower-latency positioning.

Multi-turn consistency

One successful edit is not enough. A realistic sequence might change product color, replace a background, modify a headline, and then adjust the aspect ratio.

Each step can undo an earlier decision. OpenAI’s image-generation guide covers multi-turn editing, but I would evaluate complete sequences rather than isolated requests.

The useful measurement is whether the final asset preserves all the constraints accumulated across those turns.

Reference fidelity

The 2.5 family is intended to retain distinctive subject features when changing setting, composition, or style.

That includes faces, clothing, packaging, product geometry, and visual identity. For e-commerce variants or avatar transformations, losing those details can invalidate an otherwise polished image.

The official reference-fidelity example illustrates the intended behavior: an old portrait is transformed while retaining the child’s identity and overall pose, despite substantial changes to clothing and presentation.

Official Images 2.5 reference-fidelity input.

Official Images 2.5 edited output.

I would use examples like this to define tests, not as evidence that every identity-preserving edit will work equally well.

Layouts, style, and transparency

OpenAI also reports better adherence to complex visual instructions and more accurate content involving real-world information. The generation supports transparent backgrounds, with details in the image-generation guide.

For production graphics, I would check layout constraints and transparency alongside appearance. A visually convincing result can still fail if the composition or background does not match the requested asset specification.

Latency that affects iteration

Flare’s claimed 50% lower latency is especially interesting because the published token rates did not increase.

I would measure end-to-end request time rather than assume every resolution and quality configuration gets the same improvement. Sunburst sits at a different point on the curve: slower generation in exchange for tighter control.

Same token rates does not mean the same image bill

OpenAI publishes the same token schedule for Flare, Sunburst, and Image 2:

Token category	Flare / 1M tokens	Sunburst / 1M tokens	Image 2 / 1M tokens
Text input	$5.00	$5.00	$5.00
Cached text input	$1.25	$1.25	$1.25
Image input	$8.00	$8.00	$8.00
Cached image input	$2.00	$2.00	$2.00
Image output	$30.00	$30.00	$30.00

The important distinction is between price per token, cost per request, and cost per accepted asset.

Quality, resolution, token consumption, and retry count all affect the final bill. Equal rates do not establish equal token usage. A model that needs fewer revisions may be cheaper at the workflow level even when its individual requests are not.

For a unified multi-model API, CometAPI lists both 2.5 models at $4 per million input tokens and $24 per million output tokens, 20% below the corresponding $5 text-input and $30 image-output OpenAI rates. That comparison should not be confused with OpenAI’s separate image-input or cached-input categories.

My migration spreadsheet would therefore track accepted assets, not just successful HTTP responses.

Integration surface: choose the model, then measure the result

The generation and editing endpoints described for access are:

POST /v1/images/generations
POST /v1/images/edits

The available model choices are:

gpt-image-2.5-flare
gpt-image-2.5-sunburst
gpt-image-2

The access flow is straightforward: create an API key in the provider’s dashboard, choose a model ID, and submit a generation request with the prompt, size, and quality settings. Use the edits endpoint when modifying an existing image.

After retrieving the generated image, save or display it and review latency, token usage, and quality before increasing traffic.

I would not equate that small API-level change with a low-risk behavioral migration. Changing a model identifier is easy; validating the output distribution is the work.

How I would route workloads

Workload	Starting choice	Reason
High-volume social assets	Flare	Throughput and lower latency
Rapid prototypes	Flare	Faster iteration
Visual search	Flare	Latency-sensitive use
E-commerce variations	Flare	Balance of fidelity and volume
Final advertising creative	Sunburst	Quality and edit control
Product hero imagery	Sunburst	Reference preservation and precision
Complex revisions	Sunburst	Strongest editing score in this snapshot
Text-heavy premium graphics	Sunburst	Strongest text-rendering score
Stable existing pipeline	Image 2 or staged migration	Known behavior and more mature benchmark evidence
New general-purpose integration	Flare	Practical default trade-off

A routing strategy worth testing is to generate drafts with Flare and use Sunburst for difficult final edits. I would only keep that split if evaluation shows it improves total turnaround or acceptance rate.

Sunburst is not automatically the best choice because it leads most quality categories. Extra precision has to justify the extra latency for the task.

My migration checklist

For a controlled comparison, I would hold the prompt, reference image, resolution, and quality configuration fixed wherever the controls are comparable.

Then I would collect:

• End-to-end latency
• Token consumption
• Accepted-output rate
• Number of revisions or regenerations
• Subject and product consistency
• Typography accuracy

I would include multi-turn sequences and difficult localized edits, not only clean text-to-image prompts.

Image 2’s larger benchmark sample is a legitimate reason to migrate cautiously. Existing integrations do not need to move immediately, and the dated snapshot remains useful while evaluating alternatives.

The decision metric I care about is cost per accepted image at an acceptable turnaround time. If a model costs the same per token but requires half as many regeneration attempts, the workflow economics can change substantially.

Safety and provenance still belong in the evaluation

The Images 2.5 system card reports unsafe generations reaching users in:

Model	Unsafe generations in adversarial tests
Sunburst	1.09%
Flare	1.41%
ChatGPT Images 2.0 baseline	1.64%

These results come from deliberately adversarial prompts. They are not normal-production incident rates.

OpenAI also reports continued use of C2PA provenance metadata and invisible watermarking. Better photorealism increases the potential impact of deceptive generated media, so I would keep provenance and misuse evaluation in scope rather than treating quality as an isolated concern.

Visual examples are useful test prompts, not benchmarks

The supplied official examples cover product imagery, precision editing, and style:

Official OpenAI product visual from the Images 2.0 release.

Official OpenAI precision-editing example.

Official OpenAI style example.

I would separate these demonstrations from the independent preference scores. They help identify capabilities to evaluate, but they do not replace a representative workload.

My default: Flare first, Sunburst where it earns its latency

The 2.5 upgrade targets the parts of image generation that tend to create production friction: reference preservation, localized revisions, repeated edits, typography, layouts, and iteration speed.

The early Arena results support testing both models. Sunburst has the stronger quality signal across most categories; Flare combines improved preliminary scores with OpenAI’s lower-latency claim.

For a new general-purpose integration, I would start with Flare. For precision-heavy final assets, I would compare Sunburst directly. For a stable Image 2 pipeline, I would keep the baseline until controlled evaluation demonstrates better acceptance rates, turnaround, or total cost.

The model name matters less than how reliably the full workflow produces an asset I can actually use.