Skip to content
Mixen API
RU

Chat and streaming

POST /v1/chat/completions is the main endpoint. Request and response match OpenAI, so any compatible SDK works.

Parameters

Section titled “Parameters”
Field Type Description
model string, required Catalog identifier or alias (gpt-4o, claude-opus)
messages array, required Conversation so far; the last user message is the prompt
stream boolean true streams the answer over SSE
stream_options object {"include_usage": true} appends a final chunk with token usage
temperature number As in OpenAI. Not every model takes it: GPT-5.x/5.6/6 and Claude have none — check supported_parameters in GET /models
max_tokens integer Ceiling on answer length
max_completion_tokens integer Synonym for max_tokens, for newer SDKs
reasoning_effort string Reasoning depth: off, low, medium, high, xhigh or max — the set depends on the model, see reasoning effort. For routers (typesafe/jev-router) it is a hint: the router picks the effort itself and may override it
tools array Tool descriptions in the OpenAI format; passed to the model as is — how to use
tool_choice string | object auto, none, required, or a specific tool
web_search bool | string true/"auto" — the model’s native search where available, else Exa; "native"/"exa" — explicit. Sources arrive in message.annotations; full guide — web search
response_format object {"type": "json_object"} or {"type": "json_schema", …} — guaranteed JSON — structured output
stop array Up to 4 stop sequences: generation ends as soon as one appears in the answer
seed integer Sampling seed — best-effort reproducibility
top_p number 0–1 Nucleus sampling: only the most likely tokens covering this probability mass
top_k integer Keep the K most likely tokens. Available on open-weight models
min_p number 0–1 Drop tokens below this share of the top token’s probability
frequency_penalty number −2…2 Penalty grows with how often the token already appeared
presence_penalty number −2…2 Penalty for the token appearing at all
repetition_penalty number 0–2 Penalty for repeating the prompt and the answer so far
logprobs boolean Return log probabilities of the chosen tokens
top_logprobs integer 0–20 Alternatives per position to return; requires logprobs
provider object Route selection: order, only, ignore, sort — provider routing
service_tier string flex — cheaper and slower, priority — dearer and faster; not every model has a tier
plugins array Upstream plugins as-is; for typesafe/jev-router — a custom model pool: [{"id": "jev-router", "models": ["anthropic/*", "openai/gpt-6*"]}] (fields models/allowed_models, excluded_models, cost_tier)

Models accept different parameter sets: the upstream drops what a model does not know instead of erroring. The actual list per model is in the supported_parameters field of GET /models.

Routers

Section titled “Routers”

typesafe/jev-router picks the model and reasoning effort for each request itself — the model that actually served the request arrives in the served_model response field (in streaming: a separate chunk with empty choices before the usage chunk). The field appears only when the served model differs from the requested one. Billing follows the chosen model’s rates.

Streaming

Section titled “Streaming”
stream = client.chat.completions.create(
    model="gpt-5.6-luna",
    messages=[{"role": "user", "content": "Write a haiku about deadlines"}],
    stream=True,
    stream_options={"include_usage": True},
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

The stream is text/event-stream and ends with data: [DONE]. With include_usage, a chunk carrying usage arrives just before it — a convenient place to log spend: it also carries cost / cost_rub, what the request was charged (see Tokens and charges).

The usage also carries prompt_tokens_details.cached_tokens — how many input tokens came from the context cache. See Context caching.

Image input

Section titled “Image input”

Models with capabilities.vision: true accept an image inside a message:

resp = client.chat.completions.create(
    model="gpt-5.6-terra",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What's wrong with this chart?"},
            {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KG..."}},
        ],
    }],
)

Only data: URLs (data:image/png;base64,...) are accepted, in the last user message — external HTTP links are not supported, so base64-encode the image.

File input

Section titled “File input”

Models with capabilities.file: true take up to 5 files per request. Formats: PDF, DOCX, TXT, CSV, XLSX, PPTX — the text is extracted and handed to the model as context:

{
  "role": "user",
  "content": [
    {"type": "text", "text": "Summarise the Risks section"},
    {"type": "file", "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0..."}}
  ]
}

Each file yields at most 150,000 characters; longer ones are cut with a [...truncated] marker appended. There is no OCR: an image-only PDF scan comes back empty — use documents with a text layer.

A model that cannot read files returns 400, so check the flag in the catalog first.

Audio input

Section titled “Audio input”

Models with audio input (in the catalog: capabilities.audio: true, e.g. google/gemini-4-argon) take sound as an input_audio content part:

{
  "role": "user",
  "content": [
    {"type": "text", "text": "Transcribe this recording and list the agreements"},
    {"type": "input_audio", "input_audio": {"data": "//uQx...", "format": "mp3"}}
  ]
}

The data field is raw base64 without the data:audio/...;base64, prefix (unlike images); the format is a separate field: mp3, wav, flac, pcm16. Audio tokens are billed at a separate audio-input rate (on multimodal models it can differ from the text one — see the Providers panel on the model page); repeated requests with the same audio hit the audio cache at a discount.

Reasoning effort

Section titled “Reasoning effort”

Some models let you dial reasoning depth. The accepted values arrive in capabilities.reasoning_efforts — usually low, medium, high, xhigh, max.

{"model": "claude-opus-4.8", "reasoning_effort": "high", "messages": [...]}

Two things to watch:

  • Reasoning consumes the max_tokens budget. If you expect long JSON from a model that reasons, set the ceiling generously or the answer gets cut mid-structure. Truncation is visible in the response: finish_reason: "length" (or stop_reason: "max_tokens" on /v1/messages), so you can detect it programmatically and continue.
  • Reasoning cannot always be switched off — check capabilities.can_disable_reasoning. Where it reads false, the model always reasons, and the price reflects that.
Section titled “Web search”

Models with type: search (the Perplexity Sonar family) reach the internet themselves. There is no separate endpoint or tool — just name one:

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"sonar-pro","messages":[{"role":"user","content":"What did Anthropic ship this week?"}]}'

On a scoped key these models need the search scope rather than chat.

The web_search parameter — search on any model

Section titled “The web_search parameter — search on any model”

Web search can also be enabled on any other text model with the web_search parameter:

{
  "model": "google/gemini-3.8-flash",
  "web_search": "native",
  "messages": [{"role": "user", "content": "What is the central bank USD rate today? One line with a source"}]
}

Values: true/"auto" — the model’s native search where available, else Exa; "native" — the provider’s built-in search; "exa" — neutral search. Native search exists on OpenAI, Anthropic, Google and xAI models — the catalog marks them with capabilities.native_search: true; sources arrive in message.annotations (url_citation), same as Sonar. When streaming, they come as a separate chunk carrying delta.annotations right before the final chunk with finish_reason. "native" on a model without built-in search is silently downgraded to "auto". Billing, the balance reserve and Sonar’s behaviour are covered in the Web search guide.

Search is billed per performed request (not per token): the price sits next to the model in the catalog (web_search_rub_per_request in GET /v1/models). The model decides on its own whether and how many times to search — a question like “what is a fraction” triggers none.

Other protocols

Section titled “Other protocols”

Besides /v1/chat/completions, two more chat protocols run on the same key and balance:

  • POST /v1/messages — Anthropic-compatible (Claude Code, Anthropic SDK): accepts thinking.budget_tokens, cache_control and tool-use; there is also POST /v1/messages/count_tokens. See Claude Code.
  • POST /v1/responses — the OpenAI Responses API (Codex CLI): stateless, accepts reasoning.effort. See Codex CLI.