ChatGPT Images 2.5: Flare vs Sunburst Technical Review

Quick answer

OpenAI announced ChatGPT Images 2.5 on September 8, 2026. The official description focuses on turning ideas, sketches, and reference photos into more personalized and polished images that better reflect the user's intent.

For API users, the Route Key live catalog currently exposes two OpenAI entries: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Our editorial reading is that this is a workflow split, not a simple good-versus-better ladder:

• Start with Flare for frequent generation and editing where responsiveness and iteration speed matter.
• Test Sunburst for high-value production work where precise edits, visual continuity, and reviewer acceptance matter more than the fastest first draft.

Do not choose from launch demos alone. Image models fail differently from text models: a result can look impressive while changing the wrong object, losing a person's identity, corrupting text, or drifting after the third edit. The right metric is cost and time per accepted asset, not image price or first-render latency in isolation.

This article separates OpenAI's official release statement, Route Key catalog observations, third-party launch reporting, and our engineering recommendations. Sources and catalog entries were checked on September 10, 2026.

What is officially confirmed?

OpenAI's official news feed published Introducing ChatGPT Images 2.5 on September 8, 2026. Its concise product description confirms four inputs or goals worth testing:

1. Turning written ideas into images.
2. Developing sketches into finished visuals.
3. Working from reference photos.
4. Producing more personalized, polished results that better follow intent.

Those statements establish the product release and its direction. They do not, by themselves, establish a universal latency improvement, a fixed API price, a supported size matrix, or a guarantee that every ChatGPT capability appears identically in the API.

OpenAI's currently indexed image generation guide documents the Images API, image generation and editing, image inputs, output formats, quality controls, transparency, streaming partial images, and the GPT Image model family. At the time of this review, the documentation index available to us had not yet added a complete 2.5 parameter table. Treat new model controls as version-specific until the live reference confirms them.

What Route Key currently exposes

The Route Key live model catalog returned these catalog entries on September 10, 2026:

Catalog model ID	Vendor label	Billing mode	Supported endpoint labels
gpt-image-2.5-flare	OpenAI	Per request	Image generation, OpenAI compatible
gpt-image-2.5-sunburst	OpenAI	Per request	Image generation, OpenAI compatible

This verifies that both names are routable Route Key catalog entries. A gateway catalog entry does not prove that every provider-native option is available through every compatible endpoint. Confirm the live price, group, request schema, supported sizes, output format, and account access immediately before production use.

Flare vs Sunburst: the useful working hypothesis

Launch coverage consistently describes Flare as the broad, faster default and Sunburst as the higher-control option for demanding visual work. Until OpenAI's full developer reference is indexed, treat that distinction as a hypothesis to validate rather than a permanent specification.

Start with Flare when iteration is the product

Flare is the first candidate for:

• conversational image generation;
• concept exploration and mood-board variants;
• social and editorial drafts;
• ecommerce background changes;
• rapid localization experiments;
• interactive applications where users refine a result several times.

The acceptance test is not simply whether the first image looks good. Measure whether a user can reach the intended result in fewer turns and less wall-clock time.

Start with Sunburst when correction cost dominates

Sunburst is the first candidate for:

• identity-sensitive edits;
• product and packaging consistency;
• precise object replacement or removal;
• multi-reference composition;
• campaign assets with layout constraints;
• high-resolution work that receives professional review.

A slower or more expensive request can still be the economical route if it avoids manual retouching, reshoots, or several failed generations.

Do not turn the names into policy

Some easy-looking edits are identity-sensitive. Some complex illustrations benefit more from quick exploration than maximum fidelity. Route by the consequence of failure and the expected number of iterations, not by a static content category alone.

The five evaluation suites that matter

1. Prompt adherence

Create prompts with count, position, color, material, camera, lighting, and exclusion requirements. Score each requirement separately. A single aesthetic score hides whether the model ignored the operational part of the prompt.

Useful checks include exact object count, left/right relationships, legible labels, prohibited objects, aspect ratio, and safe negative space for copy.

2. Edit locality

Ask the model to change one named element while preserving everything else. Compare the input and output for unintended changes to faces, hands, logos, text, product geometry, background objects, crop, lighting, and color grade.

The strongest editing model is often the one that changes less.

3. Identity and reference fidelity

Use authorized reference images and test people, pets, products, characters, and visual styles separately. Reviewers should score recognizability, proportions, defining features, material details, and consistency across several outputs.

Do not collapse identity preservation and style preservation into one score. A model may reproduce a palette while losing the subject.

4. Text rendering and layout

Test short labels, multilingual copy, punctuation, numbers, and mixed text plus illustration. Record exact-string accuracy, missing characters, substitutions, unwanted text, hierarchy, and whether the layout leaves the required safe zones.

For production design, correct text in the wrong place is still a failed asset.

5. Multi-turn stability

Begin with one image and request five small edits in sequence. Check whether earlier approved details survive each turn. Track the edit at which identity, composition, or typography begins to drift.

This suite matters because a conversational editor can be excellent on one edit and unreliable as a working session.

Measure accepted assets, not pretty samples

Use a review table that connects quality to operations:

Metric	What it reveals
First-pass acceptance	Baseline instruction following and visual quality
Accepted asset rate	How often the output can actually ship
Median repair turns	Hidden iteration cost
Manual retouch minutes	Cost moved from API usage to design labor
Edit-locality failure rate	Risk of damaging approved content
Exact-text success rate	Suitability for ads, UI, packaging, and localization
Time to first preview	Perceived responsiveness
End-to-end task time	Real workflow speed
Cost per accepted asset	Comparable economic outcome

Third-party launch reports cite latency improvements of up to 50%. That is a vendor-reported maximum repeated by the press, not our benchmark. Network location, size, quality, queueing, moderation, number of outputs, and edit inputs can all change observed latency.

A minimal OpenAI-compatible request

Both live catalog entries advertise an OpenAI-compatible image-generation endpoint. Keep the first integration deliberately small and add model-specific controls only after confirming the current reference:

import OpenAI from 'openai'

const client = new OpenAI({
  apiKey: process.env.ROUTE_KEY_API_KEY,
  baseURL: process.env.ROUTE_KEY_BASE_URL,
})

const result = await client.images.generate({
  model: 'gpt-image-2.5-flare',
  prompt:
    'Editorial product photograph of a brushed aluminum desk lamp, neutral studio background, soft side light, no logo, no text',
})

const imageBase64 = result.data?.[0]?.b64_json

Switch the model to gpt-image-2.5-sunburst for an A/B run while preserving the prompt and other settings. Before shipping, verify the current SDK version, authentication header, response shape, output format, size, quality, background, compression, streaming, and image-edit parameters in the Route Key integration guide and live model entry.

Never place an API key in browser code or an image-generation prompt. Proxy requests through a controlled server and apply user, budget, moderation, and rate limits there.

Production architecture: save the recipe, not only the bitmap

An accepted image should retain enough information to reproduce or audit it:

• exact model ID and provider route;
• prompt and negative constraints;
• ordered reference-image identifiers;
• size, quality, format, compression, and background settings;
• seed or generation identifier when the API exposes one;
• parent asset and edit sequence;
• moderation result and reviewer decision;
• creation time, latency, and billed usage;
• license and consent state for source assets.

Store original uploads separately from generated derivatives. Strip unnecessary metadata before public delivery, but preserve an internal audit record where policy requires it.

Cost controls for image workloads

Per-request pricing is easy to read and easy to underestimate. One user action may trigger several variants, preview generations, edits, upscales, safety retries, or fallback calls.

Set budgets at the workflow level:

accepted asset cost = all generation calls + edits + upscales + storage + delivery + reviewer time

Practical controls include:

1. Generate one or two previews before a large batch.
2. Use Flare for exploration and promote an approved direction to Sunburst only when testing supports that route.
3. Hash prompts and authorized source assets to prevent accidental duplicate jobs.
4. Cancel abandoned streaming requests when the API supports cancellation.
5. Put daily limits on users, projects, and automated campaigns.
6. Retain failure reasons so prompt and routing improvements target real waste.

Do not publish a hard-coded price inside application UI unless the billing source and update process are clear. Use the live catalog because gateway price and provider list price can change independently.

Our editorial verdict

ChatGPT Images 2.5 appears to be a workflow release more than a demo release. The important product promise is not merely prettier output; it is better translation of intent from text, sketches, and references into assets that need fewer corrections.

The Flare/Sunburst split is sensible if it gives teams an explicit speed-versus-control choice. Flare should win default traffic only when it reaches acceptable results faster after repairs are counted. Sunburst should win premium traffic only when its additional control measurably reduces rejected assets or manual editing.

The launch is real, and both API names are live in Route Key's catalog. The remaining task is disciplined measurement. Until the full 2.5 developer reference is indexed, avoid copying parameter tables or fixed prices from screenshots and community posts.

FAQ

Is ChatGPT Images 2.5 officially released?

Yes. OpenAI's official news feed published the release on September 8, 2026. Availability can vary by product, account, region, and API channel.

Are Flare and Sunburst official API model IDs?

gpt-image-2.5-flare and gpt-image-2.5-sunburst are live Route Key catalog IDs labeled as OpenAI models. At the time checked, the official developer documentation index available to us had not yet exposed a complete 2.5 model reference, so verify provider-native naming separately if you call OpenAI directly.

Which model should I use first?

Begin with Flare for responsive, iterative work and Sunburst for precision-sensitive production work. Reverse that choice whenever your own accepted-asset data supports it.

Is GPT Image 2.5 really 50% faster?

Launch reporting cites an improvement of up to 50%, but that is not a guarantee for every size, quality level, region, or edit workload. Benchmark end-to-end task time under your actual request mix.

Does it support image editing and reference photos?

OpenAI's announcement explicitly mentions sketches and reference photos, and the GPT Image API family supports image editing. Confirm the exact 2.5 edit inputs, limits, and fidelity controls in the current API reference before implementation.

Should I replace GPT Image 2 immediately?

No. Replay the same generation and multi-turn edit suite, compare accepted-asset rate, latency, cost, and reviewer time, then move traffic gradually with GPT Image 2 retained as a rollback route.

Sources and related reading

1. OpenAI, Introducing ChatGPT Images 2.5, September 8, 2026.
2. OpenAI Developers, Image generation guide.
3. OpenAI Developers, Images API reference.
4. Route Key, live model catalog, checked September 10, 2026.
5. Route Key, OpenAI-compatible integration guide.
6. Route Key, GPT-6 Astra technical review.
7. Route Key, reduce AI API costs with caching and routing.