GPT Image 2.5: Flare vs Sunburst, Pricing, API and Filters

Flare and Sunburst at a glance

Developer

OpenAI

Release date

September 8, 2026

Product name

ChatGPT Images 2.5

API models

gpt-image-2.5-flare, gpt-image-2.5-sunburst

Quality levels

low, medium, high, xhigh, max, auto

Max size

3840 px longest side (above 2560×1440 experimental)

Image output price

$30 per 1M tokens ($15 Batch)

Open weights

No, closed API model

Uncensored?

No, moderation is always on

What is GPT Image 2.5?

ChatGPT Images 2.5 is the product name; the underlying model generation is OpenAI's 2.5 image line. OpenAI's model docs list two API models, Flare and Sunburst, each with a snapshot dated 2026-09-08, and press coverage on September 9 confirmed a Tuesday launch. It is available in ChatGPT (reported for all plans), ChatGPT Work and Codex, and through the Images API and the image generation tool in the Responses API. Both models take text and image input and return images. OpenAI says the release improves detail, lighting and texture, keeps subjects more consistent across multi-turn edits and makes local edits more reliable.

Why people look for an uncensored GPT Image 2.5

OpenAI states that all prompts and generated images are filtered against its content policy. The API moderation parameter only accepts auto and low; low is less strict but still blocks content, so there is no setting that turns filtering off. Artists working on boudoir, swimwear, fantasy or other adult-themed images often hit refusals or moderation_blocked errors, which is why people search for an uncensored or NSFW mode of this model. No such official mode exists, and we do not publish ways to bypass it. Spicy AI does not run OpenAI's model; its image editor uses other models without content filters for adult (18+) content. If you came here looking for uncensored GPT Image 2.5, the honest answer is that it does not exist, and the generator below is a separate, clearly labeled alternative.

From GPT Image 2 to 2.5: what changed

GPT Image 2 arrived in April 2026 according to press reports, and the 2.5 generation followed on September 8. The most visible API change is quality: 2.5 offers six levels (low, medium, high, xhigh, max and auto), adding xhigh and max on top. In ChatGPT the update brought Sketch, poster and product templates, comments on images and shared prompts, and OpenAI claims generation is up to 50% faster and that more than 3 billion images are made each week across ChatGPT Images and the API. Those two figures are vendor claims we could not verify independently. Some outlets report that 2.0 and older DALL·E flows will be retired, but we found no official date, so treat that as unconfirmed.

gpt-image-2.5-flare vs gpt-image-2.5-sunburst

Flare and Sunburst are separate API models, not quality levels. Flare is positioned for fast, high-volume text-to-image work such as social graphics, drafts and product scenes. Sunburst is positioned for editing and fidelity, keeping details of a reference image while you change one region, and it usually takes longer. Both share the same per-token prices and both work in the Images API and the Responses API. Same price does not mean same cost: the number of output tokens can differ between the two at the same size and quality, so measure the usage field in your own requests. A practical workflow is to explore layouts with Flare, then finish the chosen image with Sunburst.

From OpenAI's model pages; check them for the latest values.

Model IDs, endpoints and parameters

Use POST /v1/images/generations to create images and POST /v1/images/edits to edit them, or call the image generation tool inside the Responses API and pass the model there. Model values must be lowercase: gpt-image-2.5-flare or gpt-image-2.5-sunburst, or a dated snapshot if you want stable behavior. Results come back as base64. Useful controls are size, quality, n for several images per call, output_format (PNG by default, JPEG for speed, WebP for smaller files), output_compression and background set to transparent, which requires PNG or WebP. OpenAI warns that complex prompts can take up to about two minutes, so set generous client timeouts. Third-party gateway docs mention 1 to 10 reference images of up to 15 MB each; OpenAI's pages did not state a limit, so treat that as unconfirmed.

Max resolution and 4K

The recommended sizes are 1024×1024, 1536×1024 and 1024×1536, and custom sizes are allowed within rules: both sides must be multiples of 16, the aspect ratio must stay between 1:3 and 3:1, the longest side can be at most 3840 pixels, and the total must fall between 655,360 and 8,294,400 pixels. That means 3840×2160 (4K UHD) is possible, but anything above 2560×1440 is labeled experimental, so expect slower jobs, more tokens and less predictable results. For most web use, generating at a recommended size and upscaling separately is cheaper.

API pricing and price per image

The 2.5 generation is billed by tokens, not by image, and OpenAI has not published a fixed per-image table for it. Both models cost the same: text input $5 per million tokens ($1.25 cached), image input $8 per million ($2 cached) and image output $30 per million, with Batch output at $15 per million. OpenAI's own example is a 1024×1024 image at low quality using 196 output tokens, about $0.00588, or roughly $0.003 through Batch. That counts output only. A rough formula is output tokens × $30/1M plus input image tokens × $8/1M plus text tokens × $5/1M. Higher quality, larger sizes, reference images and multi-turn edits all add tokens, so read the usage field instead of quoting old GPT Image 2 per-image prices.

Prices from OpenAI's model pages; they can change, so confirm before budgeting.

Is ChatGPT Images 2.5 free to use?

Reports say ChatGPT Images 2.5 reached every ChatGPT plan, including free accounts, but OpenAI has not published how many images a free user gets, and limits can change with demand. The API has no free tier for these models; every call is billed per token. If you want to try image generation at no cost, Spicy AI offers daily free generations with its own models. They are not OpenAI's model, so results and style will differ.

How to use Images 2.5 in ChatGPT and Sketch

In ChatGPT, describe the image or upload a reference, then refine it one change at a time: ask for a new background, a different light or a line of text, and keep the previous version so you can go back. Sketch lets you draw a rough layout, outfit outline or room plan with @Sketch and use it as a visual reference, which helps when a composition is easier to draw than to describe. Templates give posters and product shots a starting structure, and comments let you point at a specific area. Text rendering, layout control and character consistency can still fail, so keep headlines short, state where each line goes, and reuse the same reference image and descriptors for recurring characters.

What people make with the model and how reviewers rate it

Early hands-on coverage shows where the new model gets used most. Axios tested the new editor at launch and found it faster and better at keeping people and pets recognizable across edits, which matches the most common jobs: retouching personal photos, swapping backgrounds on product shots and turning rough sketches into posters. Tom's Guide published a set of playful creative prompts, and a prompt that turns a modern selfie into a 1980s studio portrait spread on Instagram. Reception is mixed rather than glowing: TechRadar quoted Reddit users calling the noise patterns awful and saw a gritty, sandy texture in its own tests, while some users said they could not tell it apart from version 2.0. Treat these as early impressions, not benchmarks.

• Photo edits: keep the same person or pet while changing background, outfit or light.
• Product and shop images: clean backdrops, lifestyle scenes and transparent PNG cutouts.
• Posters and social graphics: start from a template, keep headlines short, refine with comments.
• Retro and style remixes, such as turning a modern selfie into a 1980s studio portrait.
• Developers: early questions focus on token cost and how retries affect the bill.

How to write prompts: a beginner tutorial

A good prompt reads like a short brief to a photographer or illustrator. Start with the subject and what it is doing, then add where it is, how it is lit, which lens or art style you want, and finally any text that must appear. Write plain sentences instead of long tag lists, because the model follows natural language well and understands relations such as on the left or behind the counter. When a result is close, edit it rather than starting over, and describe the change in one sentence. The same five steps also work in the uncensored generator on this site.

• Step 1, subject: say who or what, with age (adult), clothing, pose and expression, for example an adult woman in a linen shirt, laughing at the camera.
• Step 2, setting: name a concrete place and time, such as a small Lisbon café at dusk, instead of a nice background.
• Step 3, light and camera: choose one light source and one lens or style, such as warm window light, 85mm portrait, shallow depth of field.
• Step 4, text and layout: put the exact words in quotes and give a position, such as headline 'OPEN LATE' at the top center.
• Step 5, edit: change one variable per turn and say what must stay, such as keep her face and pose, make it night.
• Before and after: 'woman in cafe, nice' becomes 'An adult woman reading at a window table in a Lisbon café at dusk, warm lamp light, 50mm photo, muted film colors'.

Prompt templates and a step-by-step workflow

A reliable prompt for this model names five things in order: subject, setting, lighting, camera or art style, and any on-image text with its exact position. Begin with one or two drafts at low or medium quality to find the composition, then switch to high quality and edit in small steps, changing a single variable per round and keeping the previous version. In the API, draft with Flare and finish with Sunburst. The templates below are starting points: replace the bracketed parts, and keep every person clearly adult.

• Product shot: "[product] on a [surface], [color] backdrop, soft studio light from the left, 50mm lens, gentle shadow, no text"
• Poster: "Vertical poster for [event], headline '[3 to 5 words]' at the top in bold sans-serif, date at the bottom, [style] illustration in the center"
• Edit step: "Keep the person, pose and framing exactly the same; only change the background to [new scene] and match the lighting"
• Style remix: "Turn this photo into a [decade or art style] portrait, keep the face and hairstyle, add [film grain or studio backdrop]"
• Sketch to image: "Use my @Sketch as the layout: [object] on the left, title area on the right, render as a [photo, 3D or illustration]"

Flare and Sunburst vs Nano Banana 2.1 and Seedream 5.0

Pick by task rather than by a single winner. Google's Nano Banana 2.1 is reported to be cheaper per image, supports 4K output and search grounding, and is strong at keeping characters and objects consistent. ByteDance's Seedream 5.0 is known for infographics, layer separation and low prices at 1K to 2K. OpenAI's model stands out for the ChatGPT workflow, Sketch and precise multi-turn editing. All three are closed models with their own content policies. Prices change often, so check each vendor's current page before you decide.

Competitor details are summarized from public reporting and may change.

NSFW, moderation and the moderation_blocked error

A moderation_blocked error means the safety system rejected the prompt, an input image or the output. Sending the same request again will not help and wastes time. Check whether a reference photo shows a real person, nudity or a minor, rephrase descriptive wording, and remember that moderation set to low is only a little more lenient. A timeout is a different problem: look at your client or gateway timeout first. Explicit adult imagery is outside OpenAI's policy, and we do not share bypass prompts. For legal adult content with consenting, fictional adults, Spicy AI's image editor is the separate option this site offers.

Known limits and open questions

Confirmed limits are slow complex prompts of up to about two minutes, imperfect text and layout, characters that can drift between images, and moderation on every request. Still unknown at the time of writing are the retirement date for GPT Image 2, the exact free-plan quota in ChatGPT, an official cap on reference images, the real token difference between 2.0 and 2.5, and independent benchmarks. The model cannot be downloaded or run locally because there are no public weights. It suits teams doing iterative editing inside ChatGPT or the API; it does not suit anyone who needs open weights, no filtering or a guaranteed low price per image.

Prompt tips for GPT Image 2.5-style results

1. Change one thing per round, such as background, light or outfit, and keep the earlier version.
2. Write on-image text as short lines and say exactly where each line should sit.
3. Reuse the same reference image and descriptors to keep a character consistent.
4. Do not resend a moderation_blocked prompt unchanged; rewrite it or switch tools.
5. Describe every person as an adult (18+) and never use photos of real people without consent.

Sources

• gpt-image-2.5-flare model page (OpenAI)
• gpt-image-2.5-sunburst model page (OpenAI)
• Image generation guide (OpenAI)
• OpenAI launches Images 2.5 (IT Brief)
• ChatGPT Images 2.5 generates pictures 50% faster (ExtremeTech)
• Images 2.5 and the Sketch feature (Digit)
• Hands-on with ChatGPT's new image editor (Axios)
• Reddit users on Images 2.5 noise patterns (TechRadar)
• 7 prompts tested with ChatGPT Images 2.5 (Tom's Guide)

More uncensored AI Image models

• Uncensored Midjourney
• Uncensored DALL·E
• Uncensored Stable Diffusion
• Uncensored Flux
• Uncensored Nano Banana
• Uncensored Nano Banana 2
• Uncensored Nano Banana 2.1
• Uncensored Ideogram
• Uncensored Leonardo AI
• Uncensored Seedream
• Uncensored Seedream 5.0
• Uncensored Krea
• Uncensored Kroma
• Uncensored Qwen-Image
• Uncensored Adobe Firefly
• Uncensored Text to Image
• Uncensored Z-Image Turbo
• Uncensored Z-Image
• Uncensored FLUX.2
• Uncensored FLUX.2 [dev]
• Uncensored FLUX.2 [pro]
• Uncensored GPT Image 1.5