# GPT Image 2.5 (image)

> GPT Image 2.5 makes detailed still images with exact rendered typography. It is the roasOS image model for statics that carry text: headlines, body copy, wordmarks and UI mockups. This endpoint runs fal's flare model: fast, and the default.

Status: roasOS API v1.1. Catalog `model-catalog@2026-10-06.v5`, pricing `customer-pricing@2026-11-05.v15`. The endpoint `POST /v1/image/gpt-image-2.5` is live. `POST /v1/generations` stays for one minor version as a deprecated alias (see "Today").

## Endpoint
POST https://api.roasos.com/v1/image/gpt-image-2.5
Auth: Authorization: Bearer roasos_sk_… (scope api:write). Idempotency-Key required. quote_only / quote_id / webhook_url as in https://docs.roasos.com/llms.txt.

## Good at
GPT Image 2.5 is OpenAI's image model on fal. fal describes the flare model as "OpenAI's default image model for most applications. Fast, high-quality generation with natural lighting, rich textures, and support for complex layouts including transparent backgrounds". Use it for headline typography, dense body copy, wordmarks, and UI mockups such as chat threads and dialogs. Use it with up to 8 references for a maskless edit; fal permits 16, roasOS caps at 8. It accepts `avatar_id`: the saved avatar's front and full-body shots lead the reference list and count as 2 of the 8. `avatar_outfit` requires `avatar_id`. Set `background` to `transparent` for a PNG cutout with alpha, or to `opaque` for a filled canvas; the price does not change. Do not use it for photoreal handheld product shots. Do not use it for subject continuity across many references; it blends multiple references less smoothly than Nano Banana Pro. Likeness quality on GPT Image 2.5 is not yet measured; Nano Banana Pro remains the measured identity lane. For fine detail with a longer wait use /v1/image/gpt-image-2.5-sunburst at the same price. For a masked, positional edit use /v1/image-edit/gpt-image-2.5.

## Pricing
Unit: image. Pricing keys: `gpt-image-2.5-high`, `gpt-image-2.5-low`, `gpt-image-2.5-medium`.
The catalog publishes no credit example for this model: the pricing table has no key for it, and roasOS never states a guessed price.

## Modes
- `text-to-image`
- `image-edit`

The worker derives the mode from the media fields of the request.
Provider endpoints: `openai/gpt-image-2.5/flare/edit` · `openai/gpt-image-2.5/flare/text-to-image`.

## Request fields
| field | type | required | default | rule |
|---|---|---|---|---|
| `ad_look` | enum | no | — | `designed`, `native`, `ui` |
| `ad_set` | object | no | — | `{ id, index, total }` |
| `aspect_ratio` | enum | no | `1:1` | `9:16`, `1:1`, `2:3`, `3:4`, `4:5`, `4:3`, `3:2`, `16:9` |
| `avatar_id` | string | no | — | UUID |
| `avatar_outfit` | string | no | — | ≤ 500 characters |
| `background` | enum | no | `auto` | `auto`, `transparent`, `opaque` |
| `num_images` | integer | no | `1` | 1–4 |
| `product_id` | string | no | — | UUID |
| `product_images` | string[] | no | — | ≤ 8 items; only in `image-edit` |
| `prompt` | string | yes | — | ≥ 1 character; ≤ 10,000 characters |
| `prompt_mode` | enum | no | — | `wrapped`, `exact` |
| `quality_tier` | enum | no | `high` | `low`, `medium`, `high` |
| `quote_id` | string | no | — | UUID |
| `quote_only` | boolean | no | `false` | — |
| `references` | string[] | no | — | ≤ 8 items; only in `image-edit` |
| `source_image` | string | no | — | https URL; ≤ 4,096 characters |
| `wait_timeout_seconds` | integer | no | `0` | 0–60 |
| `webhook_url` | string | no | — | https URL; ≤ 4,096 characters |

The schema is strict: a field that is not in this table returns `400 unsupported_option`. Field names are the roasOS role names, one name per role across every family.

## Media
- `avatar_id` — identity; the saved character reference sheet, with a derived portrait only where the model requires a single face or source frame
- `product_id` — product; resolves to the product image list
- `product_images` — product reference images
- `references` — style and appearance references

| role | items | max size | content types | duration |
|---|---|---|---|---|
| `product_images` | ≤ 8 | 10 MB | `image/jpeg`, `image/png`, `image/webp` | — |
| `references` | ≤ 8 | 10 MB | `image/jpeg`, `image/png`, `image/webp` | — |

- Refused: `elements`, `reference_images`, `image_url`, `end_image_url`, `scene_id`, `scene_image`, `video_urls`, `video_url`, `audio_urls`, `audio_url`, `voice_id`. A refused role returns `400 unsupported_option`.
- Every URL must be an immutable object on `https://cdn.roasos.com/…` that your workspace owns; any other URL is `400 reference_media_not_workspace_owned`. Upload with `POST /v1/uploads` (direct POST of the bytes to `put_url`, or a `source_url` import), then use the returned CDN URL.

## Prompt
Limit: 10,000 characters including the wrapper (the wrapper is added only with product media). fal accepts 32,000. No tag syntax.
- Treat body-text lines as exact strings. Put each rendered string in quotation marks.
- Render critical text at 25 % or more of the canvas height.
- State exact element counts ("EXACTLY THREE message rows").
- Keep all text and focal subjects inside the central 84 % of the canvas.
- Name the layout, the type style and the color of each text block.
- For a cutout, say "isolated on a transparent background" and set `background: "transparent"`.
- On a bad result, write a new prompt that names the fix; do not send the same payload again. Stop after two retries.

## Not exposed
- `quality`: values not exposed — `auto`, `xhigh`, `max`.
- `output_compression`: not accepted by this public endpoint.
- `image_size`: pinned to `seven fixed sizes in text-to-image, auto in an edit`.
- `output_format`: pinned to `png`.
- `sync_mode`: pinned to `false`.
- `mask_url`: pinned to `reachable only through the inpaint endpoint`.
- `references`: pinned to `8 images including the avatar's two shots; fal permits 16`.
- `prompt`: pinned to `characters 10,001–32,000`.
- `seed`: not accepted by this public endpoint.
- `negative_prompt`: not accepted by this public endpoint.

## Result
Result kind `image`. Always present: nothing. Present when the run produces it: `avatar_id`, `full_body_url`, `images`, `portrait_url`, `provider_seed`, `reference_sheet`, `url`.
```json
{ "data": { "id": "gen_…", "kind": "image", "model": "gpt-image-2.5", "status": "completed", "result": {"avatar_id":"00000000-0000-4000-8000-000000000000","full_body_url":"https://cdn.roasos.com/workspaces/ws_01/full_body_url","images":["https://cdn.roasos.com/workspaces/ws_01/images-1.png"],"portrait_url":"https://cdn.roasos.com/workspaces/ws_01/portrait_url","provider_seed":0,"reference_sheet":"https://cdn.roasos.com/workspaces/ws_01/reference_sheet","url":"https://cdn.roasos.com/workspaces/ws_01/result"} } }
```
Files live on `https://cdn.roasos.com` and are immutable; a result URL can be a media input to the next call. Poll `GET /v1/jobs/{id}/status` → `{ id, status, updated_at }`, then `GET /v1/jobs/{id}/result` (`409 run_not_completed` until the job is `completed`, `failed` or `canceled`). Statuses: `queued → processing → completed | failed | canceled`.

## Example
```bash
# 1. quote — nothing is spent
curl -X POST https://api.roasos.com/v1/image/gpt-image-2.5 \
  -H "Authorization: Bearer roasos_sk_live_…" \
  -H "Idempotency-Key: 0f1c9e2a-gpt-image-2.5-0001" \
  -H "Content-Type: application/json" \
  -d '{"quote_only":true,"prompt":"A woman holds the bottle at eye level in a bright kitchen, morning light, static camera."}'
# → 200 { "data": { "kind": "quote", "quote_id": "q_7Hk…", "credits": …, "expires_at": "…", "confirmation": { "model": "gpt-image-2.5" } } }

# 2. run — the same body and the same Idempotency-Key, quote_id instead of quote_only
curl -X POST https://api.roasos.com/v1/image/gpt-image-2.5 \
  -H "Authorization: Bearer roasos_sk_live_…" \
  -H "Idempotency-Key: 0f1c9e2a-gpt-image-2.5-0001" \
  -H "Content-Type: application/json" \
  -d '{"quote_id":"q_7Hk…","prompt":"A woman holds the bottle at eye level in a bright kitchen, morning light, static camera."}'
# → 201 { "data": { "id": "gen_…", "kind": "image", "model": "gpt-image-2.5", "status": "queued", "credits_reserved": …, "links": { "status": "/v1/jobs/gen_…/status", "result": "/v1/jobs/gen_…/result" } } }

# 3. poll
curl https://api.roasos.com/v1/jobs/gen_…/status -H "Authorization: Bearer roasos_sk_live_…"
curl https://api.roasos.com/v1/jobs/gen_…/result -H "Authorization: Bearer roasos_sk_live_…"
```

## Today
`POST /v1/image/gpt-image-2.5` is live. `quality_tier` (`low`, `medium`, `high`, default `high`) selects the price; fal's `auto`, `xhigh` and `max` are not exposed, because fal publishes no price for them. `aspect_ratio` is refused in an edit; the output follows the source image. `POST /v1/generations` with `model: "gpt-image-2.5"` still reaches the text-to-image mode for one minor version; there `resolution` carries the quality tier. The retired `/v1/image/gpt-image-2` answers 404. The roasOS agent uses this model for the transparent-cutout role.

## See also
https://docs.roasos.com/models/image/gpt-image-2.5 · https://api.roasos.com/v1/catalog/image/gpt-image-2.5 · openai/gpt-image-2.5/flare/edit · openai/gpt-image-2.5/flare/text-to-image