GPT Image 2.5, released September 8, 2026, is the top-ranked image model on Arena and Artificial Analysis, and its strongest result is editing. Sunburst leads Flare on every board. It is not the best pick for photorealism, where reviewers still prefer Nano Banana Pro, and OpenAI publishes no per-image price table. A 1024x1024 image costs about 0.6 cents at low quality and 21 cents at max.
GPT Image 2.5 review: the short verdict
This GPT Image 2.5 review covers OpenAI’s September 8, 2026 image model: what it does, where it ranks, and what an image really costs. We checked OpenAI’s launch post and developer docs, the Arena and Artificial Analysis leaderboards, and named third-party tests, all in November 2026.
We did not run our own prompt batch. Every claim below comes from OpenAI, a leaderboard or a named reviewer, with the source linked.
The short version: GPT Image 2.5 is the top-ranked image model on both major boards, and editing is where it is strongest. Photorealism is the weak spot, and the price depends heavily on which quality setting you pick.
| Area | Verdict | Why |
|---|---|---|
| Image editing | Best available | #1 on Arena Image Edit and Artificial Analysis editing |
| Text-to-image ranking | Best available | #1 and #3 on Arena, #1 and #2 on Artificial Analysis |
| Text rendering | Good, not flawless | OpenAI’s docs admit placement and clarity can still slip |
| Photorealism | Fair | Curious Refuge saw “artificial fractal-like texture” up close |
| Speed | Strong | OpenAI claims up to 50% lower latency than Images 2.0 |
| Price | Depends on quality | About 0.6 cents to 21 cents for a square image |
Pick it if you edit existing images, juggle several reference photos, or need clean layouts. Skip it for the most natural photographic base images, or when cost per image at the highest quality matters.
What is GPT Image 2.5?
GPT Image 2.5 is OpenAI’s image generation and editing model, released on September 8, 2026 and sold to consumers as ChatGPT Images 2.5. It follows GPT Image 2, which launched on April 21, 2026.
OpenAI’s launch post for ChatGPT Images 2.5, September 8, 2026
OpenAI’s headline is “sharper details, faster generation, more precise editing, and better tools for creating and sharing.” It says generation latency is up to 50% lower than Images 2.0.
Flare and Sunburst: the two API models
Developers get two models, and both take text and image inputs and output images. Per OpenAI’s launch coverage and docs:
- gpt-image-2.5-flare: the default for most apps. OpenAI aims it at social content, visual search, prototyping and high-volume generation
- gpt-image-2.5-sunburst: the slower, more precise model for premium work that needs tighter edit control, such as campaign creative and polished product images
Both add xhigh and max quality settings on top of low, medium and high. They also have a separate complexity field (low, medium or high, default high) that OpenAI says carries no extra fee.
Where you can use it
OpenAI says Images 2.5 is live in ChatGPT, ChatGPT Work and Codex on desktop, mobile and web, across all tiers. The API models are available now.
The ChatGPT side adds Sketch, where you draw a rough picture and add written directions, plus templates such as Poster and Merch, in-image comments and shareable prompts.
GPT Image 2.5 key features
GPT Image 2.5’s main upgrades are reference-based generation, precision editing and better multi-turn memory. Here is what OpenAI claims, and what the docs confirm.
- Reference-based generation: subjects from reference photos stay recognizable when the setting, style or composition changes
- Precision editing: only the requested element changes, while the subject, background, layout and brand styling stay stable
- Longer edit chains: earlier edits survive through several rounds of changes
- Big canvases: edges up to 3,840 px, an aspect ratio between 1:3 and 3:1, and 655,360 to 8,294,400 total pixels. Anything above 2560x1440 is marked experimental
- Markup-guided edits: the Image API accepts markup on the first input image for supported GPT Image 2.5 models
- Transparent backgrounds and more accurate real-world information, per OpenAI
OpenAI also lists Adobe, Runway, Manus and Higgsfield AI as launch partners, and says Adobe will bring both models to Firefly. Treat the speed numbers as vendor claims: the Manus figure of two to four times faster than GPT Image 2 is a customer measurement, not an independent test.
GPT Image 2.5 rankings: Arena and Artificial Analysis
GPT Image 2.5 Sunburst is first on every major image board we checked, with Flare in the top three. Rankings below were checked on November 6, 2026, and Arena’s snapshot is dated November 2.
Artificial Analysis text-to-image leaderboard, checked November 2026
Text-to-image
On Artificial Analysis, the 2.5 models take the top two spots:
| Rank | Model | Elo | Price per 1,000 images |
|---|---|---|---|
| 1 | GPT Image 2.5 Sunburst (max) | 1227 | $210.72 |
| 2 | GPT Image 2.5 Flare (max) | 1208 | $210.72 |
| 3 | GPT Image 2 (high) | 1188 | $211.00 |
| 4 | Nano Banana 2.1 | 1176 | $33.60 |
| 5 | MAI-Image-2.6 | 1168 | $38.91 |
The cost column is the catch. Nano Banana 2.1 scores 51 Elo points below Sunburst at about one sixth of the price.
Arena’s Text-to-Image board, last updated November 2, 2026
Arena agrees. Sunburst leads at 1427 and Flare is third at 1399, ahead of GPT Image 2 at 1382, MAI-Image-2.6 at 1331 and Nano Banana 2.1 at 1325. Arena marks both 2.5 scores as preliminary.
Image editing
Editing is the biggest gap. On Artificial Analysis’s editing board, the standings are:
| Rank | Model | Elo | Price per 1,000 images |
|---|---|---|---|
| 1 | GPT Image 2.5 Sunburst (max) | 1199 | $210.72 |
| 2 | MAI-Image-2.7-Preview | 1195 | No API available |
| 3 | GPT Image 2.5 Flare (max) | 1182 | $210.72 |
| 4 | Krea 3 (Ultra) | 1152 | $100.00 |
| 5 | Nano Banana 2.1 | 1152 | $33.60 |
Artificial Analysis image editing leaderboard, checked November 2026
The top two are within error bars of each other: Sunburst is 1199 with a margin of 11, and Microsoft’s preview model is 1195 with a margin of 14. Our MAI Image 2.7 review covers that challenger and has no API or price yet.
On Arena’s Single Image Edit board, Sunburst scores 1517 with 114,468 votes and Flare 1482. MAI-Image-2.7-Preview is fourth at 1469 and Nano Banana 2.1 is eighth at 1427.
One more thing you will see on Arena: preliminary “GPT-Image-2.6” entries. We found no OpenAI announcement of a 2.6 model, so this review covers 2.5.
Text rendering, photorealism and editing: what testers found
Independent hands-on data is thin, because the model is two months old and OpenAI published no benchmark table or competitor comparison. DataNorth’s launch coverage notes the claims are all relative to GPT Image 2.
We can’t generate images for this post, and we won’t invent a side-by-side. Here is what named sources report instead.
Editing and multi-reference work
Curious Refuge tested the Sunburst version through Magnific on September 11 against Nano Banana Pro, Nano Banana 2, Seedream 5.0 and MAI-Image-2.6. Their verdict:
- GPT Image 2.5 “was the clear winner for complex multi-reference generation”
- “For targeted image edits and compositing, GPT Image 2.5 is one of the most intelligent models we’ve tested”
- A billboard text replacement kept perspective and wrapping intact
- A product swap was “remarkably convincing” on lighting and perspective
This matches the leaderboards: editing is the lead.
Photorealism
This is where the picture changes. Curious Refuge wrote that “when you zoom in, faces and other fine details can develop an artificial fractal-like texture,” and preferred Nano Banana Pro for base photographs.
They also saw a drift problem. “With every edit, the image became slightly sharper, more saturated, and more artificial,” so they suggest one or two targeted edits per generation.
If photorealism is your whole job, read our GPT Image 2.5 vs Nano Banana 2 comparison before you pay for max quality.
Text rendering
OpenAI’s image generation guide says the model is “significantly improved” at text but “can still struggle with precise text placement and clarity.” That is the vendor admitting a limit.
Curious Refuge noted “distinctly AI-looking typography choices” on posters. Picsart’s comparison page claims roughly 99% English text accuracy, but that is a vendor page, not an independent test, so we don’t count it.
Our advice: proofread every word, price and claim in a generated image. Canva built its Craft model around text and layout, if that is your main job.
GPT Image 2.5 pricing: what an image costs
OpenAI prices GPT Image 2.5 by token, at the same rates as GPT Image 2. It does not publish a fixed per-image price for 2.5.
From OpenAI’s API pricing page, checked November 2026, for both Flare and Sunburst:
| Token type | Standard (per 1M) | Batch (per 1M) |
|---|---|---|
| Image output | $30.00 | $15.00 |
| Image input | $8.00 | $4.00 |
| Cached image input | $2.00 | $1.00 |
| Text input | $5.00 | $2.50 |
OpenAI’s model page says its older GPT Image 2 calculator “does not estimate GPT Image 2.5 token consumption,” so there is no official table to quote.
Estimated price per 1024x1024 image
The best available numbers come from tokencost.app, which rebuilt OpenAI’s calculator and multiplied output tokens by the $30 rate. Treat them as estimates, and expect input images to add cost.
| Quality | Output tokens | Estimated cost |
|---|---|---|
| low | 196 | $0.0059 |
| medium | 439 | $0.0132 |
| high | 1,756 | $0.0527 |
| xhigh | 3,122 | $0.0937 |
| max | 7,024 | $0.2107 |
The max row lines up with Artificial Analysis, which lists $210.72 per 1,000 images at max. The leaderboard scores both use max, so the ranking is the expensive setting.
That matters for like-for-like cost. GPT Image 2 (high) is $211.00 per 1,000 on the same list, so the top tier costs about the same as before. On the tokencost.app estimate above, a “high” request now costs about a quarter of what GPT Image 2’s “high” did, because the labels map to different token tiers.
GPT Image 2.5 pros
- Best editing on the boards. #1 on Artificial Analysis editing and Arena Image Edit
- Strong with references. Curious Refuge called it the clear winner for multi-reference generation
- Two models for two jobs. Flare for speed and volume, Sunburst for tighter edits
- Flexible output. Up to 3,840 px edges and ratios from 1:3 to 3:1
- Cheap at low settings. About 0.6 cents for a square low-quality image
- Easy access. It is in ChatGPT on all tiers, with Sketch and templates
GPT Image 2.5 cons
- Costly at the top. About 21 cents a square image at max, against $33.60 per 1,000 for Nano Banana 2.1
- No official price table. You have to estimate cost from tokens
- Photorealism lags. Reviewers still prefer Nano Banana Pro for natural base photos
- Edit drift. Repeated edits made images sharper and more artificial
- Text is not solved. OpenAI’s own docs warn about placement and clarity
- Preliminary scores. Arena marks the 2.5 results as preliminary, and the model is two months old
GPT Image 2.5 vs alternatives
GPT Image 2.5 tops the boards, while rivals win on price or style. Elo is from Artificial Analysis’s text-to-image board, and prices are per 1,000 images, checked November 2026.
| Model | Elo (text-to-image) | Price per 1,000 | Best for |
|---|---|---|---|
| GPT Image 2.5 Sunburst (max) | 1227 | $210.72 | Edits, references, layouts |
| GPT Image 2.5 Flare (max) | 1208 | $210.72 | Faster everyday generation |
| Nano Banana 2.1 | 1176 | $33.60 | Lower cost per image |
| MAI-Image-2.6 | 1168 | $38.91 | Low-cost generation |
For style-first work, Midjourney runs on subscriptions rather than per-image pricing, so see our Midjourney V8 review for plan costs. If you want to animate stills, the best text-to-video AI ranking compares video models.
Who should use GPT Image 2.5?
- Marketers and designers who edit: yes. Sunburst’s edit control is its main strength
- Product and ecommerce teams: yes, for swaps and reference-based shots. Test lighting on your own products
- Developers building apps: yes. Start with Flare, move to Sunburst for premium edits, and set quality to match your budget
- Photographers wanting natural base images: maybe. Compare against Nano Banana Pro first
- Cost-sensitive volume users: use low or medium quality, or look at cheaper models
- Anyone who needs flawless text: not yet. Proofread every output
Conclusion: is GPT Image 2.5 the best?
Mostly yes. Our GPT Image 2.5 review comes down to this: it leads Arena and Artificial Analysis in generation and editing, and Sunburst is the best-ranked editing model with a public API.
It is not the best at everything. Photorealism trails Nano Banana Pro in the one detailed hands-on test we found, text still needs a proofread, and the top quality setting costs about 21 cents a square image.
Start on Flare at medium quality, which is about 1.3 cents a square image, and move to Sunburst only when an edit needs tighter control.