Chat Completions

POST https://api.prisminference.com/v1/chat/completions

Use this endpoint with the OpenAI SDK or any client that supports the Chat Completions wire format. Prism processes request and response content with strict zero data retention. Inputs and outputs are never used for model training.

Request

curl -X POST "https://api.prisminference.com/v1/chat/completions" \
  -H "Authorization: Bearer $PRISM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash",
    "messages": [
      {"role": "system", "content": "You are a senior TypeScript engineer."},
      {"role": "user", "content": "Explain the bug in this function: ..."}
    ],
    "temperature": 0.2
  }'

Body parameters

Field	Type	Required	Description
`model`	string	Yes	A Prism model ID.
`messages`	array	Yes	Conversation messages in chronological order.
`stream`	boolean	No	Return Server-Sent Events when `true`. Defaults to `false`.
`max_tokens`	integer	No	Maximum number of generated tokens, including reasoning tokens.
`temperature`	number	No	Sampling temperature. Lower values are more deterministic.
`top_p`	number	No	Nucleus sampling probability.
`stop`	string or string[]	No	Stop generation when any supplied sequence is produced.
`tools`	array	No	OpenAI function definitions available to the model.
`tool_choice`	string or object	No	Control whether and which tool the model calls.
`response_format`	object	No	Request JSON mode or output that matches a JSON Schema.
`reasoning_effort`	string	No	`none`, `minimal`, `low`, `medium`, `high`, `xhigh`, or `max`. See Reasoning.
`reasoning`	object	No	OpenRouter-style controls: `enabled` and `effort`.
`stream_options`	object	No	Set `include_usage: true` to receive usage in the final stream chunk.
`conversation_id`	string	No	Stable conversation identifier for cache-aware routing. Maximum 256 characters.
`prompt_cache_key`	string	No	OpenAI-style cache key. Used as the routing key when `conversation_id` is absent.

Each message has a role and content:

Role	Purpose
`system`	Set behavior and instructions for the model.
`developer`	Add application-level instructions.
`user`	Supply a user request.
`assistant`	Include earlier assistant text or tool calls.
`tool`	Return a tool result using its `tool_call_id`.

Conversation affinity

For multi-turn requests, include a stable, opaque conversation_id in the JSON body. Use the same value for every turn in one conversation and a different value for each separate conversation. Prism uses it to keep related requests on the same inference worker so repeated prompt prefixes can be reused.

{
  "model": "gemma-4-31b",
  "conversation_id": "conversation-123",
  "messages": [
    { "role": "user", "content": "Help me plan a migration." },
    { "role": "assistant", "content": "What system are you migrating?" },
    { "role": "user", "content": "A PostgreSQL database." }
  ]
}

conversation_id controls routing only. It does not store conversation state, so include the complete message history required for each response.

Image input

deepseek-v4.1-flash accepts images in user messages using the standard OpenAI image_url content part. Pass the image as a base64 data URL:

{
  "model": "deepseek-v4.1-flash",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "Describe the bug shown here." },
        {
          "type": "image_url",
          "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." }
        }
      ]
    }
  ]
}

JPEG, PNG, GIF, and WebP data URLs are supported. Remote URLs are rejected so the inference worker never fetches an untrusted location. The optional detail field accepts auto, low, or high. Image input is supported by deepseek-v4.1-flash and gemma-4-31b. The complete JSON request, including base64 expansion, must fit within the 9.5 MB request limit.

Reasoning

DeepSeek-V4.1-Flash reasons before it answers. When the request carries no reasoning control, the provider default applies and the first content token arrives only after the model finishes thinking. Set reasoning_effort to shorten or disable that phase:

Value	DeepSeek-V4.1 tier	Behavior
`none`	`none`	No reasoning. The first token is answer text.
`minimal`	`low`	Same as `low`.
`low`	`low`	Brief reasoning for inner-loop coding and tools.
`medium`	`high`	More reasoning for multi-step implementation work.
`high`	`xhigh`	Deep reasoning for hard problems.
`xhigh`	`xhigh`	Same as `high`.
`max`	`max`	The deepest reasoning tier. Slowest and most tokens.

Gemma 4 accepts the same values; minimal runs as low, and xhigh and max run as high.

{
  "model": "deepseek-v4.1-flash",
  "messages": [
    { "role": "user", "content": "Rename this function safely: ..." }
  ],
  "reasoning_effort": "none"
}

reasoning: { "enabled": false } is equivalent to reasoning_effort: "none", and reasoning.effort accepts the same values, so OpenRouter-style clients work unchanged. When both fields are present, reasoning_effort wins unless reasoning.enabled is false. Reasoning text is returned in message.reasoning_content and, when streaming, in delta.reasoning_content chunks that precede the first delta.content. Reasoning tokens count toward max_tokens and usage.completion_tokens and are billed as output tokens. Benchmarks that time only delta.content measure thinking time as time to first token; use none or low when latency matters more than deliberation.

Response

{
  "id": "chatcmpl_01J...",
  "object": "chat.completion",
  "created": 1788998400,
  "model": "deepseek-v4.1-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The shared counter is updated without synchronization..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 42,
    "completion_tokens": 88,
    "total_tokens": 130
  }
}

finish_reason is commonly:

Value	Meaning
`stop`	The model completed normally or reached a stop sequence.
`length`	Generation reached the configured token limit.
`tool_calls`	The model requested one or more tools.
`content_filter`	Generation stopped because of a safety policy.

Streaming

Set stream: true. The response uses text/event-stream and ends with data: [DONE].

data: {"id":"chatcmpl_01J...","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}

data: {"id":"chatcmpl_01J...","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"reasoning_content":"The counter is shared"},"finish_reason":null}]}

data: {"id":"chatcmpl_01J...","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"The race"},"finish_reason":null}]}

data: {"id":"chatcmpl_01J...","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]

Tool-call arguments may span multiple chunks. Concatenate deltas by choice and tool-call index before parsing the JSON arguments.

Errors

Errors use the OpenAI-compatible envelope:

{
  "error": {
    "message": "The requested model does not exist.",
    "code": "model_not_found"
  }
}

See Errors and retries for status codes and retry guidance.

Get started

Security and privacy

API reference

Guides

Resources

Chat Completions

Request

Body parameters

Conversation affinity

Image input

Reasoning

Response

Streaming

Errors

​Request

​Body parameters

​Conversation affinity

​Image input

​Reasoning

​Response

​Streaming

​Errors

Request

Body parameters

Conversation affinity

Image input

Reasoning

Response

Streaming

Errors