OpenAI shipped ChatGPT Images 2.5 on September 8, 2026. Two days later, the question in every Slack thread and forum post is the same one: does anything I built on GPT Image 2 break, and is the new one worth moving to?
Short version — nothing breaks. GPT Image 2 still runs, still answers the same endpoints, and still costs the same per token it did in April. The 2.5 release is an upgrade you opt into, not a migration forced on you.
That makes this less of a head-to-head than a changelog. Both models come from the same lab, on the same autoregressive line that replaced DALL·E, and they read the same prompts. The useful question is narrower: which specific things changed, how much do they matter for your work, and where does the old model still win. Below is every change worth knowing, in the order they'll affect you.
Quick Answer: The 2.0 to 2.5 Diff
| What changed |
GPT Image 2 (Apr 2026) |
GPT Image 2.5 (Sep 2026) |
| Speed |
Baseline |
Up to 50% lower latency, per OpenAI |
| Quality tiers |
low, medium, high |
low, medium, high, xhigh, max |
| Model count |
One model |
Two: Flare (fast default) and Sunburst (precision) |
| Editing |
Edits often re-render the whole frame |
Targeted edits; multi-turn changes hold better |
| Reference photos |
Up to 16 images |
Same limit, better subject preservation |
| API list price |
$8 / $30 per 1M image in / out tokens |
Identical |
| Batch API |
Supported |
Not supported at launch |
| ChatGPT extras |
Standard generate and edit |
@Sketch, Templates, prompt sharing, comment edits |
Two rows do most of the decision-making. The price row means switching costs you nothing in list terms on the API. The Batch row applies only to developers calling OpenAI's API directly, and only if they were running overnight jobs at a discount.
Everything else is a quality-and-speed judgment you can settle in about twenty minutes of side-by-side testing. There's a procedure for that near the end.
Both generations sit one dropdown apart on this site — switch mid-thread any time.
Open GPT Image 2.5 →
Change 1: Speed, and What "50% Faster" Actually Means
OpenAI's headline claim is that Images 2.5 cuts generation latency by up to 50% compared with Images 2.0. On the API side, the framing is more specific: GPT-Image-2.5 Flare is described as delivering higher image quality than GPT-Image-2 at 50 percent lower latency.
Some launch partners reported larger gaps — Flare running roughly 2 to 4 times faster than GPT-Image-2 at comparable or better quality. Treat that range as vendor-reported rather than measured; it depends heavily on which quality tier each side ran, and on request size.
The part worth internalizing: "up to 50%" is a ceiling, not an average, and it applies to Flare. Sunburst goes the other direction — OpenAI describes it as taking longer to generate in exchange for precision. If you swap gpt-image-2 for gpt-image-2.5-sunburst and your p95 latency climbs, that's the model working as documented, not a regression.
For interactive products — anything where a user is watching a spinner — Flare is the meaningful change. For batch or background rendering, latency was never the bottleneck and this row shouldn't drive your decision at all.
Change 2: The Quality Ceiling Moved Up Two Notches
GPT Image 2 offered three quality settings: low, medium, high. GPT Image 2.5 keeps those and adds xhigh and max (plus auto), on both Flare and Sunburst.
This is the single most consequential change for anyone who ever hit a wall with GPT Image 2's high tier and had no next step. There is now a next step, and then another one after that.
The catch is cost. Public per-image estimates for 2.5 at 1024×1024 land around $0.006 at low, roughly $0.053 at high, and roughly $0.21 at max. Read those as estimates — the models bill per token, output token counts vary by image, and your real invoice will move around. But the shape is clear: max is roughly an order of magnitude more expensive than high, and lands about where GPT Image 2's top tier already sat.
So "2.5 costs the same as 2.0" is true on the rate card and misleading in practice. It costs the same at the same tier. If the reason you're upgrading is to use the tiers that didn't exist before, your bill goes up — you're buying something new, not the same thing cheaper.
The LM Arena numbers, with the caveat attached
At launch, the LM Arena text-to-image leaderboard put GPT-Image-2.5 Sunburst first at roughly 1421 and Flare second at roughly 1399, with GPT-Image-2 behind at about 1381. The edit arenas showed the same ordering.
The caveat matters more than the ranking. Both 2.5 entries were preliminary, scored on a few thousand votes each, while GPT-Image-2's number rests on tens of thousands. Preliminary Elo on thin vote counts moves. A ~20–40 point gap between 2.5 and 2.0 is real but modest — it is not the difference between "usable" and "unusable," and it is smaller than the gap most people assume from reading the launch coverage.
Change 3: Editing Is the Real Upgrade
If you only read one section, this one. GPT Image 2's most-complained-about behavior was that edits re-rendered rather than edited. Ask it to change a subject's jacket color and it would regenerate the scene — new lighting, slightly different face, background furniture rearranged. Ask for three sequential changes and by the third the image no longer resembled the first.
The 2.5 launch materials attack exactly those failure modes. OpenAI describes three specific improvements:
- Surgical edits. The model changes the requested element and leaves the rest of the frame intact.
- Multi-turn stability. Earlier changes survive later ones, so a chain of edits degrades less.
- Subject preservation from references. Subjects from reference photos stay recognizable across new settings, styles, and compositions — with lighting and textures described as more natural and distinctive features more likely to carry through.
Sunburst is the variant built for this. OpenAI positions it as offering an extra level of precision for detailed creative work, and it's the one to reach for when an edit has to be exact rather than fast.
Note what did not change: the input limits. The edit endpoint accepted up to 16 reference images, each under 50MB, on GPT Image 2 as well. You're not being handed more inputs in 2.5 — you're being handed a model that uses the same inputs better. That distinction matters if you were planning an architecture change around a limit that was never the constraint.
A lot of launch coverage lists 4K output and transparent backgrounds under "new in 2.5." Both are older than that, and it's worth being precise, because if you're on GPT Image 2 you may already have capabilities you think you're missing.
The size rules apply across the gpt-image-2 family — gpt-image-2, gpt-image-2.5-flare, gpt-image-2.5-sunburst, and their dated snapshots alike:
- Arbitrary resolutions, passed as a
WIDTHxHEIGHT string such as 1536x864
- Both edges must be multiples of 16
- Maximum edge 3840px
- Aspect ratio between 1:3 and 3:1
- Total pixels between 655,360 and 8,294,400
- Anything above 2560×1440 is treated as experimental
That last line is a standing caution on both models, not a 2.0-era limitation that 2.5 fixed. Test your specific 4K use case rather than assuming it renders consistently.
Transparent backgrounds followed a similar path. GPT Image 2 launched in April without alpha-channel support, then gained it through an API update in August 2026 — background="transparent" with output_format set to png or webp (JPEG can't carry alpha). The 2.5 models support the same parameter from day one. So if you upgraded specifically for transparent PNGs, check your current model first; you may already have them.
What genuinely improved here is fidelity at the top of the size range rather than the range itself — sharper faces, cleaner product surfaces, more legible typography and small print, per OpenAI's description. That's a rendering-quality claim, not a specification change.
Change 5: One Model Became Two
GPT Image 2 was a single API model with a quality dial. GPT Image 2.5 ships as two separate models, and this is a design decision you now have to make per request rather than per project.
gpt-image-2.5-flare — the default for most applications. Faster, cheaper in wall-clock terms, and per OpenAI already higher quality than GPT-Image-2 despite the speed gain. Default snapshot gpt-image-2.5-flare-2026-09-08.
gpt-image-2.5-sunburst — built for precision on demanding creative work and tight editing control, with longer generation times. Default snapshot gpt-image-2.5-sunburst-2026-09-08.
Both carry the same token rate card, both expose the same six quality settings, and both support only two endpoints: v1/images/generations and v1/images/edits.
Why the split matters: with one model, your quality-versus-speed tradeoff was one parameter and you could tune it after the fact. With two, the tradeoff is baked into routing. Product teams building anything non-trivial will end up writing a rule — Flare for first drafts and user-facing generation, Sunburst for final renders and edits that must land exactly — which is a small amount of architecture GPT Image 2 never asked for.
There's also a smaller change on the provenance side: 2.5 outputs carry C2PA metadata plus an invisible watermark. If your pipeline strips or inspects metadata, that's worth a look before you switch.
Change 6: The ChatGPT-Only Features
Several of the most visible additions are ChatGPT product features, not model capabilities. They don't reach the API at all, so if you're an API consumer you can skip this section entirely.
- @Sketch — type
@Sketch in a chat and draw directly in ChatGPT. The drawing acts as a visual reference for the final image, which is a far more direct way to specify layout than describing it in prose.
- Templates — presets for common formats: poster, merch, flyer, product photo, logo, infographic, thumbnail, ad.
- Prompt sharing — share a prompt so someone else can run it with their own photos and details.
- Comment-based edits — place a comment on a region of a generated image to direct an edit there.
Availability is broad: the update rolls out to ChatGPT, ChatGPT Work, and Codex users across all tiers including free, on desktop, mobile, and web, subject to rate limits. OpenAI notes that more than 3 billion images are created weekly across ChatGPT Images and the API GPT-Image models — context for why the free tier gets rate-limited rather than excluded.
If your workflow lives in the ChatGPT app, @Sketch is probably a bigger day-to-day change than anything in the model itself.
What It Costs on This Site
The API list price is identical between the two models, but per-image cost is where the decision actually lands for most people. On this site both models run from the same credit balance, so you can compare directly:
| Resolution |
GPT Image 2 |
GPT Image 2.5 Draft |
GPT Image 2.5 Standard |
GPT Image 2.5 Ultra |
| 1K |
3 credits |
4 credits |
6 credits |
22 credits |
| 2K |
5 credits |
5 credits |
8 credits |
24 credits |
| 4K |
8 credits |
6 credits |
14 credits |
42 credits |
Image-to-image on 2.5 costs the same as text-to-image at each tier. There's also a GPT Image 2 Stable variant at 4 / 8 / 24 credits for 1K / 2K / 4K, which trades cost for consistency on the older model.
Read the table honestly and one pattern stands out: GPT Image 2 is cheaper at every resolution than the comparable 2.5 tier, and at 1K and 2K it undercuts even 2.5 Draft. The exception is 4K, where 2.5 Draft at 6 credits actually beats GPT Image 2 at 8 — so if high-resolution drafting is your main job, the upgrade pays for itself immediately.
Credits come from the same pool either way. There's a free weekly check-in worth 30 credits, subscription plans starting at $11.90 for 100 credits, a $29.90 / 300-credit Standard plan that includes a commercial license, and one-time packs from $9.90 for 80 credits. Full breakdown on the pricing page.
Reasons to Stay on GPT Image 2
Upgrade guides usually end with "upgrade." This one shouldn't, because there are three concrete reasons to sit still, and two of them are structural rather than preference.
Per-image cost at volume. On this site, 1K generation runs 3 credits on GPT Image 2 versus 4 on 2.5 Draft and 6 on Standard. At a hundred images that's a rounding error; at ten thousand it's the difference between one plan tier and the next. Volume work where "good" beats "best" — thumbnails, variant testing, internal mockups, placeholder art — has no reason to move.
Batch API, if you call OpenAI directly. A note for API users only: gpt-image-2 supports the Batch endpoint (v1/batch); neither 2.5 model does at launch, and no batch pricing discount exists for them yet. This has no bearing on generating images in the browser — it only matters if your own pipeline hits OpenAI's API and depends on batch throughput or pricing.
Prompts and workflows you've already tuned. GPT Image 2 has been in production since April 2026 and has accumulated far more community prompt knowledge than a model that's two days old. Its LM Arena score rests on tens of thousands of votes, which is a reasonable proxy for "thoroughly characterized." If you have a prompt library that reliably produces what your brand needs, the 20–40 Elo points on offer may not survive re-tuning all of it.
The honest summary: switch for editing quality, for the xhigh and max ceiling, or for interactive speed. Don't switch out of version anxiety.
How to Decide in Twenty Minutes
Rankings are averages over other people's prompts. Yours are not average. The only test that answers your question is the one you run on your own work, and running it is cheap:
- Pick three prompts you actually use — one that works well today, one that's borderline, one that GPT Image 2 has never handled properly.
- Run all three on the GPT Image 2 tool page at your normal resolution. Save the outputs.
- Run the identical prompts on the GPT Image 2.5 tool page in Draft, then in Standard. Same wording, same resolution.
- Take the best result from each model and run three sequential edits on it — change one element, then another, then a third. This is where the two models separate most visibly, and where 2.5's multi-turn claim either holds up on your content or doesn't.
- Compare the credit cost of the path that produced the result you'd actually ship.
Step 4 is the one people skip and the one that matters. Single-shot generation quality between these two models is close enough that reasonable people disagree. Multi-turn editing behavior is where the gap is visible without squinting.
If you'd rather start from prompts that are already known to work, the prompt library is a reasonable source of test material for both models.
FAQ
Is GPT Image 2.5 better than GPT Image 2?
On OpenAI's own claims and LM Arena's launch rankings, yes — but by a smaller margin than the version number suggests. The clearest wins are editing precision, multi-turn stability, and speed on Flare. For straightforward single-shot generation at standard resolutions, many users won't see a decisive difference.
Did the API price go up with GPT Image 2.5?
No. Both gpt-image-2.5-flare and gpt-image-2.5-sunburst use the same token rate card as gpt-image-2: $5 per 1M text input tokens ($1.25 cached), $8 per 1M image input tokens ($2 cached), and $30 per 1M image output tokens. What can raise your bill is the new xhigh and max tiers, which consume far more output tokens per image.
What's the difference between Flare and Sunburst?
Flare is the fast default for most applications, delivering higher quality than GPT-Image-2 at roughly half the latency. Sunburst trades generation time for extra precision on detailed creative work and exacting edits. They share pricing, quality tiers, and endpoints — the difference is entirely speed versus precision.
Do my GPT Image 2 prompts still work on 2.5?
Yes. Both models sit on the same autoregressive multimodal line and read prompts the same way, so nothing needs rewriting to run. Expect to re-tune rather than rewrite: 2.5 follows editing instructions more literally, so prompts that over-specified to compensate for 2.0 re-rendering everything may now be doing unnecessary work.
Can I use GPT Image 2.5 for free?
In ChatGPT, Images 2.5 reached all tiers including free at launch, subject to rate limits. On this site, the free weekly check-in gives 30 credits, which covers roughly seven 1K Draft generations or five Standard ones — enough to run the comparison above before deciding whether to buy anything.
Pick one prompt you know well, run it on both GPT Image 2 and GPT Image 2.5, and let your own output settle it. Twenty minutes beats twenty leaderboard screenshots.
Sources
- GPT-Image-2.5 Flare model reference — OpenAI — verified 200: token pricing, six quality tiers including xhigh and max, 2026-09-08 snapshot, Batch endpoint listed as not supported.
- GPT-Image-2.5 Sunburst model reference — OpenAI — verified 200: identical rate card and tiers, inpainting support, batch unsupported, rate limits by tier.
- GPT-Image-2 model reference — OpenAI — verified 200:
gpt-image-2-2026-04-21 snapshot and Batch (v1/batch) listed as a supported endpoint on the older model.
- Introducing ChatGPT Images 2.5 — OpenAI — the launch announcement (up to 50% lower latency, Flare and Sunburst positioning, subject preservation, 3B images per week). Page blocks non-browser requests; confirmed via search indexing rather than direct fetch.
- ChatGPT Images 2.5: Faster, more precise, but not the same for everyone — The Decoder — verified 200: per-image cost estimates at 1024×1024, preliminary LM Arena scores and vote counts, C2PA and invisible watermarking.
- OpenAI Releases ChatGPT Images 2.5 With Sketch and Two New API Models — Unite.AI — verified 200: @Sketch, Templates, prompt sharing, comment-based edits, and rollout across ChatGPT, ChatGPT Work, and Codex tiers.
