Skip to content
Mixen API
RU

Models and pricing

The catalog comes from GET /v1/models. No key required — you can call it from your frontend to show users a current list.

curl https://api.mixen.ai/v1/models

The catalog holds 170+ models across every modality. Both the line-up and the prices change, so read them from the API rather than from this page.

What it costs

Section titled “What it costs”

The cheapest models in each modality, so you can size up the numbers without leaving the docs.

Text

Section titled “Text”
ModelInputOutputCache readCache write
deepseek/deepseek-v4.1-flash$0.1880$0.7500$0.0040—
z-ai/glm-5.3-flash$0.1880$0.6250$0.0380—
openai/gpt-6-luna$0.1250$0.6250$0.0120$0.1560
deepseek/deepseek-v4-flash$0.1140$0.2280$0.0220—
openai/gpt-5.6-luna$0.2500$1.50$0.0250$0.3120
xiaomi/mimo-v2.6-flash$0.1750$0.3500$0.0040—
anthropic/claude-haiku-5.5$0.1200$0.6000$0.0120—
google/gemini-3.8-flash$1.03$5.16$0.1020$0.0520
z-ai/glm-5.3$0.5000$10.00$0.4990—
tencent/hy4-preview$1.04$3.13$0.0520—
nvidia/nemotron-3-ultra-550b-a55b$0.7500$3.00$0.1500—
openai/gpt-6.1-sol$2.50$12.50$0.1250$3.13

per 1M tokens. A dash in the cache columns means the model has no prompt caching. Snapshot of 2026-10-16; 170 models in the catalog. Live values come from GET /v1/models.

Images

Section titled “Images”
ModelPrice
gemini-2.5-flash-image$0.0300
gemini-3.1-flash-image-preview$0.0600
gemini-3.1-flash-lite-image$0.0300
gemini-3-pro-image-preview$0.0900
google/gemini-nano-banana-2.1$0.0675
recraft-v4.1$0.0525
recraft-v4.1-utility$0.0525
recraft-v4.1-vector$0.1200
recraft-v4.1-flash$0.0105
gpt-image-2.5-flare$0.0450
gpt-image-2.5-sunburst$0.0450
gpt-image-2$0.0450

per image. Snapshot of 2026-10-16; 170 models in the catalog. Live values come from GET /v1/models.

Video

Section titled “Video”
ModelPrice
alibaba/wan-3.0$0.0750
alibaba/wan-3.0-prime$0.1020
black-forest-labs/flux-3-video$0.2550
minimax/hailuo-3$0.1950
minimax/hailuo-3-max$0.0750
x-ai/grok-imagine-video-1.5$0.1200
alibaba/happyhorse-1.1$0.1482
x-ai/grok-imagine-video-1.5-lite$0.0300
kwaivgi/kling-v3.0-pro$0.2520
kwaivgi/kling-v3.0-std$0.1890
bytedance/seedance-2.5$0.1542
bytedance/seedance-2.0$0.1009

per second. Snapshot of 2026-10-16; 170 models in the catalog. Live values come from GET /v1/models.

Response shape

Section titled “Response shape”
{
  "object": "list",
  "data": [
    {
      "id": "anthropic/claude-opus-4.8",
      "object": "model",
      "owned_by": "mixen",
      "type": "text",
      "display_name": "Claude Opus 4.8",
      "pricing": {
        "prompt_rub": 534.8,
        "completion_rub": 2673.9,
        "prompt_usd": 6.0,
        "completion_usd": 30.0,
        "cache_read_rub": 53.5,
        "cache_write_rub": 668.5,
        "currency_note": "per 1M tokens"
      },
      "capabilities": {
        "vision": true,
        "file": true,
        "reasoning": true,
        "reasoning_efforts": ["low", "medium", "high", "xhigh", "max"],
        "can_disable_reasoning": true
      },
      "supported_parameters": ["max_tokens", "reasoning", "response_format", "stop", "tools"]
    }
  ]
}

The type field tells you which endpoint accepts the model:

type Endpoint Priced per
text /chat/completions 1M tokens
search /chat/completions 1M tokens
image /images/generations, /images/edits per image; MAI — per 1M tokens by actual usage
video /videos second of video
music /audio/music request
tts /audio/speech 1000 input characters; Seed — per audio minute, Gemini TTS — per audio token
stt /audio/transcriptions per audio minute/second/hour or per 1M tokens — units in the catalog
embeddings /embeddings 1M input tokens

The unit is always spelled out in pricing.currency_note, so a client never has to infer it from the model type. Text models with prompt caching also carry cache_read_rub and cache_write_rub — the context-cache read and write prices per 1M tokens (the numbers in the example above are an illustration; take live values from the catalog).

Two more optional fields on text models: reasoning_rub_per_1m — the provider’s separate rate for thinking tokens (not present on all models: when it matches the output price, the field is absent) and web_search_rub_per_request — the price of one performed web-search request (the web_search parameter). Both are computed by the same formula as the main prices and appear only when present.

supported_parameters lists the request fields this model actually accepts. The sets differ: neither GPT-5.x/5.6/6 nor Claude take temperature, while top_k and min_p exist only on open-weight models. The upstream silently drops what it does not know instead of erroring, so check the list in code rather than by noticing a strange answer. An empty list means we have no data on the model yet, not that it accepts nothing.

Image and video models carry extra fields with the values they accept: aspect_ratios, resolutions, sizes, seconds, and references (whether the model takes reference images). Read them from the catalog instead of guessing — anything else returns 400. Speech models carry a voices list.

Naming a model in a request

Section titled “Naming a model in a request”

The model field accepts:

  • the full catalog identifier — openai/gpt-5.6-luna;
  • the short suffix — gpt-5.6-luna;
  • a familiar name from another vendor — gpt-4o, claude-opus, dall-e-3, gpt-image-1, whisper-1, tts-1, text-embedding-3-small.

Those aliases exist so a drop-in swap works: code written for OpenAI runs unchanged and lands on a comparable Mixen model. If nothing matches, the response is a 404 with code model_not_found and the message “The model ‘X’ does not exist or is not available.”

Deprecated models

Section titled “Deprecated models”

Vendors retire models. So that a hard-coded model name does not break your integration, every catalog entry carries three fields:

Field Meaning
deprecated true — the retirement date has passed; false — the model works as usual
deprecated_at retirement date and time in ISO 8601; a future date is an announcement, you have time to migrate; null — no retirement scheduled
deprecation_redirect_to id of the replacement model that will serve requests after retirement; null — no replacement

What happens to a request for such a model:

  • before the date — served as before, but the response carries the header X-Mixen-Deprecation-Notice: <date>; redirect_to=<replacement>;
  • after the date, replacement exists — the request is served by the replacement, no code change needed. The model field in the response still echoes what you asked for; the substitution is reported by the headers X-Mixen-Deprecated-Model, X-Mixen-Model-Redirect, X-Mixen-Deprecated-At (on every response, streaming included) and by a warnings array in a non-streaming body:
"warnings": [{
  "code": "model_deprecated",
  "message": "Model 'anthropic/claude-haiku-4.5' was deprecated on 2026-09-14; the request was served by 'anthropic/claude-sonnet-5'. Update your integration.",
  "requested_model": "anthropic/claude-haiku-4.5",
  "model": "anthropic/claude-sonnet-5",
  "deprecated_at": "2026-09-14T00:00:00+00:00"
}]
  • after the date, no replacement — 410 with code model_deprecated.

A redirected request is billed at the replacement’s rate — the model that actually served it; the amount is in the response’s usage.cost. Retired models with a replacement stay in GET /v1/models with deprecated: true, so you can see where you will be routed. Redirects work in the text protocols (/v1/chat/completions, /v1/responses, /v1/messages, /v1beta); media models carry the fields too, but a retired media model answers 404.

Billing

Section titled “Billing”

You are charged exactly what the catalog shows at the time of the request — no hidden multipliers. Rouble prices are not fixed: prompt_rub can differ week to week while prompt_usd stays put, so read a fresh value from GET /v1/models before you bill a client, rather than a cached one.

Charges come off your Mixen balance, the same one the bot and the web account use.

Human-readable catalog

Section titled “Human-readable catalog”

A browsable catalog with filters, descriptions and sample output lives at api.mixen.ai/models. It also shows which models accept images, take files, and support reasoning effort.

Model guides

Section titled “Model guides”

Dedicated guides with parameters, examples and limitations — for popular models: