GPT Image 2.5: Flare / Sunburst Selection & Hands-On

Updated: 2026-09-09. Product names, API model IDs, ChatGPT / Work / Codex rollout scope, and billing follow OpenAI’s introduction, ChatGPT, and OpenAI Platform for that day. IDs below reflect documentation clues around 2026-09-08—re-check once live in your account.

Introduction

ChatGPT Images 2.5 (often searched and spoken as GPT Image 2.5) is OpenAI’s next-generation image stack rolling out around 2026-09-08 across ChatGPT, Work, Codex, and related surfaces, with distinct API model IDs. Versus the prior flagship GPT Image 2 (gpt-image-2 / the Images 2.0 generation), official messaging highlights sharper detail, stronger subject hold from references, and more reliable multi-round edits. On latency, the vendor claims up to about 50% lower than Images 2.0 (a vendor claim—validate with your account’s measured P50/P95 and that day’s docs).

This guide does not stack review adjectives. It locks three rules into team docs: product display name vs API ID, Flare (default-fast) vs Sunburst (polish-stable), and a reproducible generate → edit → ship loop. Family context: What is GPT Image?. Prior flagship: GPT Image 2 complete guide.

What this guide solves

• Separate the “ChatGPT Images 2.5” product name from API identities such as gpt-image-2.5-flare / gpt-image-2.5-sunburst so contracts and logs do not record the wrong ID
• Decide Flare vs Sunburst with a table—not by guessing from a UI nickname
• Know when GPT Image 2.5 is worth switching to, and when to pin the old ID for A/B
• Build a reproducible path: acceptance brief → composition → single-point edit → reference constraints → final check
• Catch subject drift, multi-round collapse, and unverified model identity before delivery

Product name vs API name: align identity first

When you write standards, treat these four as one set—not a single nickname:

Record	Suggested wording (re-verify that day)
Product / UI label	ChatGPT Images 2.5 (may appear in batches on ChatGPT, Work, Codex, etc.)
Spoken / search common	GPT Image 2.5, Images 2.5
API default-fast	gpt-image-2.5-flare
API polish-stable	gpt-image-2.5-sunburst

Easy traps:

1. The name you pick in ChatGPT ≠ the model string on the Platform bill. The UI may only show “Images 2.5” or a shorter marketing label; automation and reconciliation must use the docs model ID.
2. 2.5 is not one ID. At least Flare and Sunburst differ; if the team only writes “use 2.5,” latency and edit behavior will not reproduce.
3. Rollout is staggered. Official notes describe rollout across ChatGPT / Work / Codex paths. Missing it in your account today does not mean the model is gone—or that a third-party entry has caught up. Trust that day’s picker and Platform list.
4. Older articles may still treat GPT Image 2 as the flagship default. Read the GPT Image 2 guide before deciding whether new delivery moves to 2.5, and to which sub-ID.

API calls, parameters, and log fields: GPT Image API guide. Hub navigation: GPT Image guides hub.

Flare vs Sunburst: how to choose

From public clues around 2026-09-08, 2.5 exposes at least two API identities (names per docs that day):

Dimension	Flare (gpt-image-2.5-flare)	Sunburst (gpt-image-2.5-sunburst)
Role	Default-fast: explore composition, batch drafts, low-round trials	Polish-precise: reference subject hold, complex multi-round local edits
Latency feel	Usually better for fast iteration (vendor also stresses latency gains vs the 2.0 generation)	Usually slower; trades for steadier edits and detail
Best use	Rounds 0–1 to find direction; pipelines that “generate candidates then human-pick”	Pre-final polish, brand/product consistency, when single-spot edits still drift
Switch when	Composition unset, quota-sensitive, need several draws to compare	Composition locked; issues are “match reference,” “change one region,” “survive multi-round”
In team docs	Log as default explore ID + verification date	Log as polish ID; do not mix with Flare in the same A/B batch without labels

Quick decisions:

1. New chat, direction unclear, need several compositions → start with Flare.
2. Composition locked; next is “match reference face/product geometry” or “change one of material / text / background” with high consistency needs → switch to Sunburst (or the matching polish tier in a supporting UI).
3. Both in one project → log model, prompt summary, reference hash, and date; do not reconstruct “which 2.5 tier” from memory later.
4. Price, quota, quality/size surcharges → do not hard-code; see Platform and account pages that day.

The “up to ~50% lower latency vs Images 2.0” line is marketing: useful as a selection motive, not as an external SLA. Ship against your environment’s measured latency.

Versus GPT Image 2: upgrade or pin?

Dimension	GPT Image 2 (gpt-image-2, etc.)	GPT Image 2.5 (this guide)
Generation role	Prior flagship workhorse; many pipelines may still pin here	New product line; API further splits Flare / Sunburst
Official emphasis	Stronger instruction follow, edits, and on-image text vs the 1.5 generation	Sharper detail, better reference subject hold, steadier multi-round edits; latency claimed down vs the 2.0 generation
Selection meaning	Historical batch parity, compliance-pinned IDs, pre-migration baseline	Evaluate first for new delivery; switch Flare/Sunburst by task
Migration note	Keep the old ID to run same-prompt cross-generation A/B	Do not assume the UI label “2.5” equals one specific API sub-ID

Recommendations:

• New projects, external key art, strong reference consistency or multi-round polish → prefer trying 2.5, then pick Flare or Sunburst from the table above.
• Pipelines, contracts, or snapshots pinned to gpt-image-2 → keep 2 as baseline; open a parallel branch on 2.5; change the default ID only after acceptance.
• Just want faster drafts → confirm 2.5 Flare is enabled on your account; if not, keep the Image 2 workflow—do not stall waiting.

Deeper Image 2 patterns (composition rounds, text strategy, quality) still apply via the GPT Image 2 complete guide—the skeleton fits 2.5; what changes is model identity and when you switch to the polish tier.

Generate + edit workflow (reproducible)

Goal: not a novel-length prompt—validate one thing per round. On 2.5, also log: this round used Flare or Sunburst.

Step 0: Write the acceptance brief (before prompting)

Five lines in notes before ChatGPT / Work / Codex or the API:

1. Use: hero / poster key art / ecommerce main / slide illustration
2. Canvas: aspect or target size tier (e.g. 16:9), title margin if needed
3. Subject and count: how many objects, facing, whether a reference must match
4. Bans: no watermark, no garbled text, no extra limbs, no unauthorized brands
5. Model identity: Flare explore / Sunburst polish + verification date

Round 1: Composition only (prefer Flare)

Prompt order: use → subject → framing/lens → light & color → material/style → on-image text (optional later) → bans. Keep style adjectives light.

Use: 16:9 blog hero with right-side margin for a title.
Subject: one matte ceramic desktop speaker, centered slightly left.
Lens: eye level, ~50mm, soft background blur.
Light: window soft light; cool gray and warm wood.
Style: clean product photography—no illustration look.
Ban: watermarks, frames, extra cables, warped buttons, any on-image text.

Accept only: correct subject, correct aspect, enough margin, no broken structure. Wrong composition → full redraw—do not stack local edits on a bad frame. Absolute beginners: Getting started: first image.

Round 2: Change one variable (after lock, consider Sunburst)

State what stays fixed, then the change:

Keep the same lens, subject pose, palette, and background.
Do only one thing: change the speaker surface from matte ceramic to brushed aluminum.
Do not change composition; do not add text or logos.

One advertised 2.5 edge vs 2 is more reliable multi-round edits—still only if you change one variable per round. Changing light, material, text, and background together makes failures undiagnosable (model drift vs conflicting instructions).

If round 2+ is “match reference” or local polish still drifts, switch this round to Sunburst and log why.

Round 3: Add references (when consistency matters)

Upload one or more references you have rights to use, and state what each controls:

Image 1: product geometry and button layout (must match).
Image 2: brand primary colors (palette only—do not copy packaging copy).
Generate: same product on a dark desktop, 16:9, still no on-image text.

Official messaging says 2.5 is stronger at reference subject hold—that is not a free pass: product shots still need Logo, proportion, and structure checks against source files. Do not request living artists’ styles, and do not upload unauthorized assets.

Round 4: Text, size / quality, and final check

• Keep on-image text short, with exact copy in quotes; for zero-tolerance spelling prefer “no text in image + layout in design tools.”
• Edit text in its own round: Keep composition and subject; only change the title to: "…".
• Only after composition and detail pass, pick target size and quality options supported that day. Higher quality does not fix a bad composition or wrong spelling.
• Deeper prompting: Prompt playbook.

Done checklist

• Entry and model identity logged (UI label / gpt-image-2.5-flare or gpt-image-2.5-sunburst / date); not confused with GPT Image 2
• Explore vs polish rounds used the right Flare / Sunburst; switch reasons are in the log
• Aspect matches use; title/crop safe margins are enough
• Subject count, hands/edges, reflections, and shadows have no obvious collapse
• Multi-round edits change one variable each time, with “keep fixed” items stated
• Reference roles documented; usage rights recorded; subject checked against references
• On-image text spell-checked, or switched to “no text + post typesetting”
• No surprise logos, watermarks, signatures, or unauthorized brands
• Final-checked at target size and chosen quality; thumbnail and full-res both reviewed
• Price, quota, and “~50% latency” marketing lines are not frozen into external SLAs; docs and measured results that day win

Fast path

• GPT Image 2.5 Flare (third-party): Open Flare generate
• GPT Image 2.5 Sunburst (third-party): Open Sunburst generate
• Model page (third-party): GPT Image 2.5
• Official ChatGPT: chatgpt.com
• OpenAI Platform: platform.openai.com
• Official introduction: Introducing ChatGPT Images 2.5

Third-party portals (including image25.xin) are not OpenAI’s account system; read that entry’s privacy and data terms before uploading commercial assets. Flare / Sunburst availability, quotas, and billing follow that site that day—do not assume third-party equals official capability that day.

FAQ

Are ChatGPT Images 2.5 and API gpt-image-2.5-* the same thing?

Same product generation, but product display names and API model IDs need not match character-for-character. In ChatGPT / Work / Codex you may only see “Images 2.5”; on Platform you explicitly pick gpt-image-2.5-flare or gpt-image-2.5-sunburst (per docs list that day). In team docs, record entry, UI label, docs ID, and verification date together.

Should I default to Flare or Sunburst?

Explore composition and draw candidates → Flare; after lock, reference hold and multi-round polish → Sunburst. If the UI shows a single “2.5” option, follow the UI notes or hidden model mapping—do not assume it is Flare.

We already know GPT Image 2 well—should we cut over to 2.5 everywhere now?

No. For new delivery, open an A/B branch: same prompt, same references, run 2 and 2.5 (with sub-ID noted) and score subject hold, multi-round stability, and latency against your acceptance brief. When pipelines pin the old ID, keep the baseline until migration passes. Details: GPT Image 2 guide.

Can we promise “up to ~50% lower latency” externally?

Do not write it as a hard SLA. It is a vendor marketing claim relative to Images 2.0, affected by size, quality, queue, region, and whether you use the polish tier. Externally, promise only “Platform / account pages that day plus the measured window both parties agree.”

What are the prices and rate limits?

No hard-coded unit prices or quotas here. See platform.openai.com and your account pages that day; third-party entries have their own plans and caps.

Multi-round edits keep collapsing—what now?

Return to a branch with correct composition, or start a new chat with a good frame as reference; for polish stages use Sunburst and strictly change one thing per round. If the prompt is bloated, strip style adjectives first and keep only constraints and bans.

Official resources

• Introducing ChatGPT Images 2.5
• ChatGPT
• OpenAI Platform
• OpenAI.com

Further reading

• GPT Image guides hub
• What is GPT Image? Family explainer
• GPT Image 2 complete guide
• GPT Image API guide
• Getting started: first image
• GPT Image prompt playbook

Summary

ChatGPT Images 2.5 / GPT Image 2.5 is the next image line after GPT Image 2: align product name and API sub-IDs first, then switch between Flare (default-fast) and Sunburst (polish-stable) by task. Drive prompts from an acceptance brief, lock composition before single-point edits, prefer Sunburst for reference constraints and multi-round polish, and treat latency/price claims as docs-plus-measurement—not frozen promises. Run the checklist before delivery, and write “entry + UI label + model ID + date” into team standards.