Images

Two endpoints: POST /v1/images/generations for text-to-image and POST /v1/images/edits for working from a reference.

Generation

Python

import base64

img = client.images.generate(
    model="flux-2-pro",
    prompt="Portrait of a ginger cat in a spacesuit, studio lighting",
    # aspect_ratio/resolution — расширения Mixen: OpenAI SDK такие kwargs
    # не принимает, поэтому их передают через extra_body
    extra_body={"aspect_ratio": "2:3", "resolution": "2k"},
)
with open("cat.png", "wb") as f:
    f.write(base64.b64decode(img.data[0].b64_json))

TypeScript

const img = await client.images.generate({
  model: 'flux-2-pro',
  prompt: 'Portrait of a ginger cat in a spacesuit, studio lighting',
  aspect_ratio: '2:3',
  resolution: '2k',
})
await Bun.write('cat.png', Buffer.from(img.data[0].b64_json, 'base64'))

cURL

curl https://api.mixen.ai/v1/images/generations \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "flux-2-pro",
    "prompt": "Portrait of a ginger cat in a spacesuit, studio lighting",
    "aspect_ratio": "2:3",
    "resolution": "2k"
  }'

Field	Description
model	An image model identifier, or the dall-e-3 / gpt-image-1 aliases
prompt	What to draw
n	How many images, 1 to 4
size	Explicit WxH
aspect_ratio	Aspect ratio such as 16:9; defaults to 1:1
resolution	1k, 2k or 4k; defaults to 1k
quality	low or medium — only on models that expose it
background	transparent | opaque | auto — GPT Image 2.5 only; transparent requires output_format of png or webp
output_format	png | jpeg | webp — GPT Image 2.5 only; defaults to png
response_format	Accepted for compatibility and ignored: the answer always carries b64_json, url is never returned

size and the aspect_ratio + resolution pair say the same thing two ways. Pick one.

Accepted values differ per model and arrive in the catalog as sizes, aspect_ratios and resolutions. Anything outside those lists returns 400.

Price depends on resolution: on most models 4k costs noticeably more than 1k. Check per_image_rub in the catalog.

Prompt length cap

Prompts are length-capped, the limit differs per model and arrives in the catalog as prompt_max_chars. Exceeding it returns 400 before any billing:

Model	Maximum
GPT Image 2.5 (Flare/Sunburst)	20,000 characters
GPT Image 2	32,000 characters
MAI-Image 2.5 / 2.6 (all tiers)	20,000 characters
Grok Imagine Image 2.0	8,000 characters
Qwen Image 3, Flux 2	no hard rejection — long text is silently truncated by the vendor
Nano Banana, Seedream	limit not published

The caps were measured by live upstream rejections on a deliberately huge prompt (verified 2026-09-10) plus official docs; always check the current value in the catalog’s prompt_max_chars.

GPT Image 2.5: transparent backgrounds and up to 16 references

The gpt-image-2.5-flare (speed tier) and gpt-image-2.5-sunburst (precision tier) series adds:

• Transparent background — background: "transparent" cuts the object out with no backdrop; the output format must be png or webp (JPEG has no alpha). auto lets the model decide from the prompt, opaque forces a filled background.
• Up to 16 reference images on /v1/images/edits — combine products, style references and characters.
• Prompts up to 20,000 characters (prompt_max_chars in the catalog) — paste briefs whole.
• Aspect ratios from 1:3 to 3:1 — including ultra-wide 3:1 and vertical 1:3, up to 4K.

curl https://api.mixen.ai/v1/images/generations \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2.5-flare",
    "prompt": "Mixen AI logo, neon lightning bolt, monochrome",
    "aspect_ratio": "1:1",
    "background": "transparent",
    "output_format": "png"
  }'

The catalog declares the knobs as background and output_formats; sending them to any other model returns 400.

Editing from a reference

POST /v1/images/edits takes multipart/form-data: one or more images in the image field, plus the same parameters as generation.

curl https://api.mixen.ai/v1/images/edits \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -F model=gpt-image-2.5-sunburst \
  -F image=@cat.png \
  -F prompt="Dress the cat as an astronaut, keep the background"

Whether a model accepts reference images at all is the boolean references flag in the catalog.

Reference billing: for models the upstream meters with per-image SKUs (Qwen Image 3, Grok Imagine and the like), every input image is added to the charge — the base per generated image plus a per-reference rate (the “Input / image” column in the providers table on the showcase). Models without that rate stay flat-priced.

One practical note: when transferring style, a model inherits the reference’s colour palette along with its composition. If you want a different palette, say so explicitly in the prompt.

Nothing async here

Both endpoints answer synchronously with the finished image. Generation takes anywhere from a few seconds to half a minute — set your HTTP client timeout to at least 120 seconds, or you will abandon a request you have already paid for.

Price

Price before you run

POST /v1/images/price accepts the same body as POST /v1/images/generations and returns the amount without generating or charging anything. It uses the same formula as the actual charge (resolution × n), so the quote matches the bill; validation is shared too — invalid parameters return the same 400/404 as a real run.

The exception is token-metered models (microsoft/mai-image-*): the upstream bills by actual tokens, so the charge follows the generation’s usage receipt — the quote stays a close estimate (same as duration with Seed Audio).

{
  "object": "price",
  "model": "gpt-image-1",
  "model_id": "gpt-image-2",
  "price": {"rub": 12.0069, "usd": 0.13},
  "per_unit": {"rub": 4.0023, "unit": "image"},
  "quantity": 3,
  "params": {"aspect_ratio": "16x9", "resolution": "2k", "size": null, "quality": null, "n": 3}
}

The rouble amount uses the live rate — quote right before you run, don’t cache it.