GPT Image 2.5 is OpenAI’s September 8, 2026 image generation and editing family. It arrived as ChatGPT Images 2.5 for everyone in ChatGPT, and as two API models: GPT-Image-2.5 Flare, the fast default, and GPT-Image-2.5 Sunburst, the precision flagship. OpenAI says people already create more than 3 billion images a week with its GPT-Image models.
The headline is the ranking. Within a month of launch, Sunburst and Flare sit at #1 and #2 on Arena’s text-to-image leaderboard, ahead of GPT Image 2, Nano Banana 2.1, Grok Imagine Image 2.0 and Microsoft’s MAI-Image-2.6. The catch: both entries are still flagged preliminary, and the per-token prices did not change — so what you actually pay per image depends on the quality setting you choose.
Two models, one family
Flare (gpt-image-2.5-flare) | Sunburst (gpt-image-2.5-sunburst) | |
|---|---|---|
| Role | Small, fast model — the default for most applications | Base model — precision for premium, detail-heavy work |
| OpenAI’s pitch | Brings the quality, editing and speed gains of Images 2.5 (OpenAI cites up to 50% lower latency than Images 2.0) | Higher precision for detailed creative work, with longer generation time |
| Docs rating | Performance “Higher”, speed “Very fast” | Performance “Highest”, speed “Medium” |
| Best for | Social and creator content, product experiences, visual search, rapid prototyping, high-volume generation | Campaign-ready creative, polished product imagery, edit-heavy workflows |
| Snapshot | gpt-image-2.5-flare-2026-09-08 | gpt-image-2.5-sunburst-2026-09-08 |
OpenAI’s prompting guide suggests a practical workflow: start with Flare when speed matters, start with Sunburst when quality requirements are demanding, then test the other model on the same prompts and keep whichever meets your bar at lower latency. Note the slightly different wording — the launch post says Flare beats GPT Image 2 on quality, while the prompting guide calls it “comparable” to GPT Image 2 and Sunburst the one with higher quality.
What’s new compared with GPT Image 2
- More faithful reference-photo work. Subjects stay recognizable when moved into new scenes, styles or compositions, with more natural light and texture.
- Targeted edits. Change a product, background or line of text without disturbing the rest of the frame — the use case OpenAI stresses for API customers.
- Steadier multi-turn edits. Earlier changes are more likely to survive later ones, instead of image quality eroding over a long conversation.
- Harder briefs. Better handling of real-world information in images, complex layouts and transparent backgrounds, and more accurate visual styles.
- New quality tiers.
xhighandmaxjoinlow,medium,highandauto.
In ChatGPT, the launch also added Sketch (draw directly in the chat and use it as a visual reference — type @Sketch), templates for formats such as posters and merchandise, comments pinned directly on an image for targeted edits, and an option to share the prompt behind an image so others can remix it with their own photos.
Pricing
Both models use the same token rates as GPT Image 2:
| Token type | Per 1M tokens |
|---|---|
| Text input | $5.00 (cached $1.25) |
| Image input | $8.00 (cached $2.00) |
| Image output | $30.00 Standard · $15.00 Batch |
Text output is not billed because the models output images. OpenAI notes that equal token rates do not mean equal cost per image — token consumption varies by model and quality setting, and its older GPT Image 2 calculator does not estimate GPT Image 2.5 usage. Read the usage field in API responses for your own prompts, sizes and quality settings.
For a rough reference, Artificial Analysis lists about $210.7 per 1,000 images (about $0.21 each) for both Flare and Sunburst at max quality, versus $211.0 for GPT Image 2 at high — and $33.6 per 1,000 for Nano Banana 2.1. Use quality: "low" for drafts and compare higher settings only for final assets.
How it ranks
| Leaderboard (as of early October 2026) | Sunburst | Flare | GPT Image 2 | Nano Banana 2.1 |
|---|---|---|---|---|
| Arena text-to-image (Elo) | 1425 (#1, preliminary) | 1398 (#2, preliminary) | 1383 (#3) | 1328 (#6, preliminary) |
| Artificial Analysis AA-Image-T2I v2.0 (Elo) | 1199 (max) | 1192 (max) | 1172 (high) | 1159 |
Arena and Artificial Analysis use different scales, so compare within a column, not across rows. Vote counts matter too: on Arena’s text-to-image board Sunburst has about 17,800 votes and Flare about 16,900, against roughly 93,000 for GPT Image 2 and 5,300 for Nano Banana 2.1 — expect the preliminary numbers to move.
Editing tells the same story. On Artificial Analysis’s AA-Image-Editing v2.0 board, Sunburst (max) scores 1184 and Flare (max) 1165, versus 1140 for both MAI-Image-2.6 and Nano Banana 2.1. Artificial Analysis’s per-image prices are measured at 1024x1024 with each model’s default settings.
API details
- Endpoints: the Image API (
/v1/images/generations,/v1/images/edits) and the Responses API image-generation tool (set the tool’smodelto either ID). - Quality:
low,medium,high,xhigh,max,auto(defaultauto). - Sizes: recommended
1024x1024,1536x1024,1024x1536; customWIDTHxHEIGHTup to 3840 px per edge and 8,294,400 pixels total (4K), with anything above2560x1440experimental. - Formats: PNG (default), JPEG, WebP; transparent backgrounds with PNG or WebP. JPEG is faster than PNG if latency matters.
- Rate limits: Build tier 250,000 TPM and 20 images per minute; Launch 3,000,000 TPM and 150; Grow 8,000,000 TPM and 250. No free tier.
- Provenance: C2PA metadata and an invisible watermark, plus prompt and image checks on generation.
How it compares
- vs GPT Image 2: same token prices, higher Arena Elo, new
xhigh/maxtiers and sharper edit control. Move to Flare first if GPT Image 2 already meets your quality bar and you want lower latency. - vs Nano Banana 2.1: GPT Image 2.5 wins on crowd-voted rankings; Nano Banana 2.1 wins on price (about $0.034 per 1K image) and Google-ecosystem fit.
- vs Midjourney v8.1: Midjourney is still the pick when the image must look art-directed; GPT Image 2.5 is the pick for literal instruction following and edit chains, and it has an API.
- vs FLUX.2: FLUX.2 [dev] can be self-hosted; GPT Image 2.5 cannot.
Limitations
- Preliminary rankings. The leaderboard numbers are early and will move as votes accumulate.
- Latency on hard prompts. OpenAI says complex prompts may take up to 2 minutes.
- Known weak spots. Precise text placement, consistent recurring characters across separate generations and layout-sensitive compositions can still fail.
- Cost transparency. There is no published per-image price; use your own usage data.
For other models in this category, see the image models directory and our Best AI Image Generator ranking.
Best for
- Precision editing — changing one product, background or line of copy while keeping the rest of the image intact
- Production campaign creative, product imagery and brand-consistent variants from reference assets
- High-volume, latency-sensitive generation inside apps (Flare)
- Infographics, UI concepts and slide visuals that must follow a detailed brief
Pros & cons
Strengths
- Sunburst and Flare hold
- Two tiers on one API — Flare for fast everyday generation, Sunburst for precision edits and premium creative work — with the same per-token prices as GPT Image 2
- New xhigh and max quality settings, custom sizes up to 3840 px on a side (4K total pixels), and transparent PNG/WebP output on both models
- OpenAI reports more precise edits that leave untouched details alone, better subject preservation from reference photos and steadier multi-turn edits
Limitations
- Per-image cost is not a fixed price — it depends on size, quality and the model's token usage, and max quality is expensive (about $0.21 per image on Artificial Analysis's default test)
- Rankings are still marked preliminary, with far fewer votes than GPT Image 2 (about 17,000 vs 93,000 on the text-to-image board)
- Complex prompts can take up to 2 minutes, and OpenAI still lists text placement, character consistency and layout-sensitive composition as weak spots
- Closed and proprietary — no weights, and the reference-photo strengths come with OpenAI's content filters and C2PA/watermark provenance marks
Frequently asked questions
When was GPT Image 2.5 released?
OpenAI launched it on September 8, 2026, together with ChatGPT Images 2.5. It rolled out the same day to ChatGPT, ChatGPT Work and Codex users on all tiers, and the two API models, gpt-image-2.5-flare and gpt-image-2.5-sunburst, were available immediately.
What is the difference between GPT-Image-2.5 Flare and Sunburst?
Flare is the small, fast model and the default for most applications. OpenAI says Images 2.5 cuts latency by up to 50% versus Images 2.0, though its prompting guide describes Flare's quality as comparable to GPT Image 2. Sunburst is the base model, built for higher precision on detailed creative work, with longer generation times.
How much does GPT Image 2.5 cost?
Both models bill tokens at the same rates as GPT Image 2 — $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens ($15 with Batch). Because token use varies by model, size and quality, measure the usage field on your own prompts instead of assuming one price per image.
Is GPT Image 2.5 better than Nano Banana 2.1?
On crowd-voted rankings, yes so far. Arena lists Sunburst at Elo 1425 and Flare at 1398 versus 1328 for Nano Banana 2.1, and Artificial Analysis shows the same order. Nano Banana 2.1 is far cheaper per image, though, at $0.0336 for 1K on Google's pricing page versus roughly $0.21 per image at GPT Image 2.5's max quality.
Can GPT Image 2.5 make transparent backgrounds?
Yes. Set background to transparent and request PNG or WebP output; both Flare and Sunburst support it.
What are the image size limits?
Both models accept custom WIDTHxHEIGHT sizes where each side is a multiple of 16, the aspect ratio is between 1:3 and 3:1, no edge exceeds 3840 px and the total is between 655,360 and 8,294,400 pixels (4K). OpenAI marks anything above 2560x1440 as experimental.