GPT Image 2.5 Is Out: Flare vs Sunburst, and Which One I'd Actually Ship

OpenAI shipped GPT Image 2.5 on September 8, 2026. In ChatGPT it's branded ChatGPT Images 2.5; in the API it arrived as two models — gpt-image-2.5-flare and gpt-image-2.5-sunburst — each with a 2026-09-08 dated snapshot. There is no gpt-image-2.5 alias, so every request has to name a variant.

Both tiers are now live on NanoBananaTool, and you pick between them with one toggle instead of reading two sets of API docs.

TL;DR

Question	Where it stands
Is it released?	Yes — September 8, 2026, ChatGPT and API the same day
What are the API models?	gpt-image-2.5-flare and gpt-image-2.5-sunburst
Which is the default?	Flare, per OpenAI: higher quality than GPT Image 2 at up to 50% lower latency
When do I use Sunburst?	Edit-heavy work where precision beats turnaround
Did prices jump?	No — OpenAI lists the same standard token rates as gpt-image-2
Is GPT Image 2 dead?	No. It's still listed with no announced retirement date
Can I use it today without an API key?	Yes — GPT Image 2.5 on NanoBananaTool

What actually shipped

Three things landed together, and it's worth keeping them separate.

In ChatGPT. Images 2.5 rolled out to ChatGPT, ChatGPT Work, and Codex across desktop, mobile, and web, alongside app-level tools: Sketch (draw a rough reference by hand), templates, image comments for pointing at a region, and prompt sharing. Those are product features, not API parameters.

In the API. Two models rather than one, split by workload instead of by generation. Both take text and image input, both do generation and editing, and both expose quality settings from low through medium, high, and two new tiers above it — xhigh and max — with auto as the default.

Everywhere else. Hosted routes lit up within a day or two, which is how most teams will touch it first. On our side, NanoBananaTool exposes Flare and Sunburst with 1K/2K/4K output and nine aspect ratios.

Flare vs Sunburst: how I actually choose

Job	My pick
Social posts, thumbnails, bulk exploration	Flare
Product retouch where one change must not disturb the frame	Sunburst
Campaign key visual with a headline	Sunburst, then QA the type
Twenty variations to find a direction	Flare, then finish the winner on Sunburst

OpenAI's own framing is simple: Flare is the default for most applications, Sunburst adds "an extra level of precision for detailed creative work with longer generation times." Community Arena scoring (still marked preliminary) put Sunburst slightly ahead of Flare on text-to-image, single-image edits, and text rendering, with both above GPT Image 2. Treat that as directional, not as a benchmark you can plan a budget around.

My working rule: explore on Flare, finish on Sunburst. Paying Sunburst's latency on drafts you throw away is the easiest way to make 2.5 feel slower than the model it replaced.

What changed versus GPT Image 2

OpenAI's claims, in its own words:

• Latency. Up to 50% lower than Images 2.0, with Flare as the fast one. "Up to" is doing work in that sentence — complex prompts can still take a while.
• Editing. More reliable instruction-following across multiple turns, changing what you asked about and leaving the rest alone. One community test ran nine consecutive edits of the same subject and reported the face, texture, and background surviving intact.
• Fidelity. More natural lighting, richer textures, better preservation of subjects from reference photos.
• Headroom. xhigh and max quality above high, plus sizes past the old 2048 ceiling — up to 3840 px on the long edge, dimensions in multiples of 16, aspect ratio no wider than 3:1, and a total pixel budget between 655,360 and 8,294,400. Transparent output is requested through a background parameter.

What did not change is just as useful:

• Token pricing. OpenAI lists the same standard rates as gpt-image-2. The "twice as expensive" takes going around compare against gpt-image-2's Batch pricing, and 2.5 has no Batch entry yet.
• GPT Image 2's status. Still a current model, no deprecation date. Keep it as your frozen baseline while you compare.
• The known weak spots. Exact spelling in small labels, pixel-perfect grid art, and precise composition still need a human pass. An early test found 2.5 couldn't hold grid-aligned pixel art the way a specialist model does.

One footnote worth knowing before you try to audit your own outputs: an independent report noted image metadata still identifies results as 2.0, so you can't use metadata to confirm which version produced a file.

Using GPT Image 2.5 without touching an API

This is the part most GPT Image 2.5 pages get wrong. They ask you to decide text-to-image or image-to-image before you've even written a prompt — a question about their plumbing, not about your work.

On NanoBananaTool there's a single page:

1. Pick Flare or Sunburst.
2. Write the prompt.
3. Optionally drop in up to 16 reference images.

If you uploaded references, it's an edit. If you didn't, it's a generation. We route to the right endpoint behind the scenes, and the credit cost shown on the button updates live as you change tier and resolution — so a 4K Sunburst frame never surprises you at checkout.

How I'd evaluate it this week

If you run anything at volume, don't migrate on vibes:

1. Freeze 30–100 representative prompts, source images, and pass/fail rules, with GPT Image 2 outputs as the baseline.
2. Score Flare and Sunburst on first-pass rate, cost per acceptable image, and p50/p95 latency — not on the prettiest single result.
3. Test your real edit chains, not one-shot generations. Multi-turn consistency is the headline claim, so it's the thing to verify.
4. Check text and labels at 100% zoom before anything goes near a campaign.

FAQ

Is GPT Image 2.5 the same as ChatGPT Images 2.5? Same release, different names. ChatGPT Images 2.5 is the product inside ChatGPT; GPT Image 2.5 is the API family, split into Flare and Sunburst.

Is there a gpt-image-2.5 model ID? No. You name Flare or Sunburst explicitly, optionally pinning the 2026-09-08 snapshot.

Did GPT Image 2.5 get more expensive? OpenAI's standard token rates match gpt-image-2. Per-image cost still depends on size and quality, so measure cost per usable image rather than comparing sticker rates.

Should I move off GPT Image 2 now? Start with Flare for everyday work, move edit-heavy jobs to Sunburst after a paired test, and keep GPT Image 2 frozen as the comparison until your own numbers say otherwise.

Where do I try it? GPT Image 2.5 on NanoBananaTool — no API key, no mode switching. For the previous generation, GPT Image 2 is still here, and GPT Image 2 vs DALL·E plus the GPT Image 2 prompt guide still apply to 2.5 prompts.