ChatCompletionRequest schema

Model identifier (e.g., deepseek-ai/DeepSeek-V4-Pro).

The conversation history. When the last message in the conversation is from the assistant, the model answers with a new assistant message. On some models this needs a short user message: "Continue." is then added, with a notice in _devup.warnings. A model that cannot answer such a conversation returns HTTP 400 conversation_must_end_with_user before any charge.

maxItems: 512

Maximum tokens to generate in the response. The largest accepted value is the model's own max_output_tokens in GET /v1/models; a larger value (for max_tokens or max_completion_tokens) returns HTTP 400 invalid_max_tokens. When neither is sent, the reply is limited to the model's default output length, which is at most its max_output_tokens. Reasoning models spend part of this budget thinking before they answer. With images, audio or video in the request, set it high enough — 2000 or more is a safe start — or the reply can come back empty with finish_reason "length".

minimum: 1

Sampling temperature (0–2).

minimum: 0; maximum: 2

Nucleus sampling probability.

minimum: 0; maximum: 1

If true, returns an SSE stream of partial completions.

default: false

Stop sequence(s) to end generation.

Penalizes repeated tokens based on frequency.

Penalizes tokens based on presence in the text so far.

A list of tools (functions) the model may call.

Controls which tool is called by the model.

How much the model reasons before answering, on models that reason on request (for example low, medium, high). Reasoning tokens are billed as output tokens.

{"enabled": true} asks for reasoning. On models that take reasoning_effort, it is applied as reasoning_effort: "medium" with a notice in _devup.warnings; when reasoning_effort is also sent, reasoning_effort applies. {"enabled": false} does not turn reasoning off on those models; a notice says so.

Response format constraint (e.g., {"type":"json_object"}). On a model that returns JSON inside a Markdown code fence, the fence around the whole reply is removed, with or without streaming, and the notice json_fence_removed is returned. On a model that does not apply json_schema itself but enforces tool calls, a json_schema request without streaming is answered through a forced tool call and returned as message.content (notice structured_output_via_tool). A reply that does not match the schema is tried once more at no cost; if it still does not match, strict: true returns 400 structured_output_mismatch (not billed), otherwise the JSON is returned with the notice structured_output_not_matched.

Request Priority or Flex processing on supported models. Omit for the standard tier. On models without the requested tier, the request is processed on the standard tier. See https://docs.devupai.com/docs/service-tiers

Allowed values: ["priority","flex"]

Identifies a cached prompt prefix. Send the same key on following requests to read the cache.

Writes the prompt prefix to the cache for the given retention time. Send it only on the request that writes the cache; every request that includes it writes again and is billed at the write price. Not allowed together with service_tier.