Is GPT Image 2.5 Uncensored? Flare and Sunburst NSFW Tests

GPT Image 2.5 gives creators a tempting choice: Flare for fast image work, Sunburst for more demanding generation and editing. If your real requirement is adult-oriented character imagery, however, the first choice is not speed versus quality. It is whether either model supports the content you need.

GPT Image 2.5 is filtered. Neither Flare nor Sunburst is an uncensored NSFW model. Both refused our September 2026 generation and editing probes, including milder wardrobe changes. Choose another foundation for a workflow that depends on NSFW output; this does not rule out every clothed portrait or romantic scene.

For a practical shortlist:

• Public promotional images: compare Flare and Sunburst on speed, editing and rendering quality.
• A mature wardrobe edit: the recorded refusals include limited edits, so neither variant is a dependable choice for this requirement.
• Higher-intensity NSFW images: start with an adult-capable generation or editing model; max quality and 4k are rendering settings, not a content-permission upgrade.

What happened in the Flare and Sunburst NSFW tests

Generation and editing used adult subjects; edits started from a clothed portrait.

Each variant refused all seven requests. The separate fine-art request used 1k, low quality and one PNG output, with 16:9 on Flare and 3:2 on Sunburst; both refusals cost $0.

Each variant refused four generation requests, including fine art, and three source-based edits.

Does moderation low make GPT Image 2.5 uncensored?

No. OpenAI's image-generation guide states that prompts and generated images are filtered. Its native moderation options are auto and low; the latter is less restrictive filtering, not disabled filtering. The guide also distinguishes blocks originating from inputs from blocks originating from outputs.

A native API screenshot showing moderation: low therefore demonstrates a filtering setting, not an uncensored mode. The recorded GPT image refusals also occurred under a low-moderation testing condition.

There is also a product distinction: the current SpicyAPI tasks for Flare and Sunburst do not expose a moderation field. Generation offers prompt, shape, resolution, quality, output count and file settings. Editing adds source images, a mask, background treatment and input fidelity. Sending an extra native API field is not a supported way to change the hosted task.

For a legitimate public-facing request, distinguish a content refusal from an upload, quota or server error. A technical error may need a corrected input or a later retry; raising image quality does not address a content refusal.

Sunburst is not the adult version of Flare

OpenAI positions Flare for fast everyday generation and Sunburst for more demanding image generation and editing. This is a capability distinction. Neither model page presents the variant as an uncensored adult-content edition.

Sunburst offers no NSFW advantage in these tests. For an accepted SFW task, the Flare versus Sunburst workflow guide compares editing controls and costs.

Flare controls with 1k, medium quality and a 4:5 frame. The price comparison uses a square frame.

Match an online example to your task

A clothed portrait, a background edit and an L3 Strong request are different jobs. Compare the exact variant, input task and requested content before treating an attractive online result as a reason to choose the model.

What you pay, and whether refusals are charged

Refused requests are not charged on either variant. The fine-art refusals cost $0.

Successful work has a separate cost. The following are the public output components for a square 1k image, identical for Flare and Sunburst:

These rows are not universal prices for every frame. Changing resolution or aspect ratio changes the output component. Long prompts add a charge for input tokens beyond the first 100, once per request. Edits also add a fixed source-image charge for each reference on every output.

That last detail matters when asking for variations. Four edited outputs with two source images each incur the reference input components for all four outputs; they do not pay for the references only once. Use the complete quote before submitting that edit.

A square 1k, medium-quality edit with one output totals $0.0207664 with one reference, or $0.0305968 with two references, for either variant. The one-reference total is $0.010936 output plus $0.0098304 input. Four separate one-reference edits at the same settings would total $0.0830656 if billed. These totals exclude any chargeable prompt tokens beyond the first 100.

Square 1k, one output and a prompt of at most 100 tokens. Medium-quality quotes rise from $0.010936 without references to $0.0207664 with one and $0.0305968 with two; source fees repeat per output.

The low entry price is useful for accepted SFW image work. It does not make either variant a cost-effective NSFW generator when the required result is refused.

Where GPT Image 2.5 can still help an adult creator

An adult project often needs ordinary public material too: a logo, an event card, a clothed fictional-character introduction or an image for a payment page. Those are sensible tasks to consider for GPT Image 2.5. They do not require promising stronger NSFW output.

For example, Flare text to image can be considered for a public character-introduction card. Start with one output at medium quality and the intended shape. Example starting prompt:

Create a photographic introduction card for two fictional adults, both aged 35, in a quiet apartment lounge. One wears an opaque navy shirt and trousers; the other wears an opaque burgundy evening dress. They stand beside each other with relaxed expressions. Leave clear wall space at the left for a short name label. Warm window light, natural proportions, no additional people and no text.

For the card, keep both characters’ complete heads and the label space in frame. The earlier clothed-couple refusal means this proposed scene may still be rejected.

For a logo or a simple non-character card, avoid unnecessary portrait references. For an approved portrait edit, use the original source and name the change precisely; adding more images increases the editing cost and can introduce conflicting people or clothing. If exact original face pixels are required, conventional compositing may be more suitable than another generative edit.

Choose an NSFW alternative by the job, not the label alone

If you need new adult-oriented images from text, Z-Image Spicy text to image is listed as unrestricted and starts at $0.01235. It is a text-generation task for images; it does not accept a reference image and therefore is not a direct replacement for editing the same approved face.

If you need an existing fictional adult character edited, Seedream 5.0 Pro editing is a reference-based alternative, listed as unrestricted with a starting output price of $0.036 at 1k. Use it when the job requires a source image; an edit can still return a softened version of the requested change.

Nano Banana 2 is not a clean substitute for an unrestricted editor. It accepted some milder generation requests in the same historical comparison but refused a corresponding lingerie edit and an L3 Strong generation brief. A completed text-only refusal is billed on Nano Banana 2. Its mixed record can be useful for milder creator imagery, but it should not be sold as a universal answer to a GPT refusal.

For the full comparison of that earlier family and Google's models, see the separate GPT Image 2 content test.

Use Flare or Sunburst when you want their image abilities within an accepted public-facing brief. Choose a different, task-appropriate foundation when the product depends on stronger NSFW images. Keep subjects clearly adult and fictional or properly authorized; an unrestricted listing does not supply consent for transforming a real person's photograph.

Frequently asked questions

Is Sunburst less censored than Flare?

The recorded tests provide no NSFW advantage for Sunburst: both variants refused all seven probes. Choose Sunburst for its image capabilities on an accepted task, not to solve a Flare content refusal.

Does moderation low turn off image filtering?

No. OpenAI describes it as less restrictive filtering. It does not disable filtering, and the SpicyAPI Flare and Sunburst tasks do not expose that native field.

Can an image edit be refused even when the source is fully clothed?

Yes. Both variants rejected the recorded wardrobe changes on a clothed source, including the limited jacket-over-camisole change. The source and the requested transformation both matter.

Will a refused request use my credits?

Refused requests are not charged on these two tasks. Successful image work uses the complete quote, including source-image fees for edits.

Which replacement should I use for NSFW work?

Use a text-to-image alternative such as Z-Image Spicy for a fresh composition. Use a reference editor such as Seedream 5.0 Pro when you need to carry an approved adult character into a new image. A cheaper generation task without source inputs cannot replace character editing.