POST /v1/chat/completions

OpenAI Chat Completions-compatible endpoint. Coding agents, IDE integrations, and any code built against the openai SDK need only change the baseURL — request shape, response shape, and streaming chunks are OpenAI’s exact wire format. Requires the ai:chat scope. This surface is stateless: it never persists conversation content and does not accept a conversation_id. Audit events are still written for every request. If you need stored conversations, run management, or PHI entity metadata, use the native POST /v1/ai/chat.

Migrating from OpenAI

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://api.usehasp.com/v1',  // ← only this line changes
  apiKey: 'hasp_key_live_...',
});

const completion = await client.chat.completions.create({
  model: 'gpt-5.4',            // a HASP model id — see "Model selection" below
  messages: [{ role: 'user', content: 'Hello' }],
});

Request

POST https://api.usehasp.com/v1/chat/completions
Authorization: Bearer hasp_key_live_...
Content-Type: application/json

Body parameters

Parameter	Type	Required	Description
`model`	string	Yes	A HASP model id (see Model selection). Not the provider’s public model name.
`messages`	array	Yes	Conversation turns. Each has a `role` of `system`, `user`, `assistant`, or `tool`.
`messages[].content`	string	—	Message text. Required for `tool` messages.
`messages[].tool_call_id`	string	—	Required on `tool` messages; correlates the result to the assistant’s tool call.
`messages[].tool_calls`	array	—	Only valid on `assistant` messages; replays a prior tool call on the next turn.
`tools`	array	No	Function-tool definitions: `{"type": "function", "function": {...}}`.
`tool_choice`	string \| object	No	Standard OpenAI tool-choice control.
`max_tokens`	integer	No	Max output tokens (1–32000). Defaults to your org’s per-request cap.
`stream`	boolean	No	Stream via SSE. Default `false`.
`temperature`	number	No	0–2.
`top_p`	number	No	0–1. Nucleus sampling.
`stop`	string \| array	No	A stop sequence, or an array of them.
`user`	string	No	Your end-user identifier. Recorded as request metadata.

Never silently dropped. temperature, top_p, stop, and tool_choice are applied to the model request. If the model’s provider cannot honor one you set, the request fails with 400 UNSUPPORTED_PARAMETER naming it. n, frequency_penalty, and presence_penalty have no equivalent in HASP’s inference request: their neutral values (n: 1, 0, 0) are accepted so default SDK payloads work, and any other value is rejected with 400 UNSUPPORTED_PARAMETER. Unlike the native /v1/ai/chat endpoint, stream defaults to false here, matching OpenAI’s default.

Model selection

The model field takes a HASP registry model id, not the provider’s public name. The same identifiers work on every HASP surface:

Model ID	Provider	Relative cost	Access
`claude-sonnet-4-6`	Anthropic	1.0×	Standard
`claude-haiku-4-5`	Anthropic	0.3×	Standard
`claude-opus-4-6`	Anthropic	1.7×	Premium
`claude-opus-4-7`	Anthropic	1.7×	Premium
`gpt-5.5`	OpenAI	1.7×	Premium
`gpt-5.5-pro`	OpenAI	10.0×	Premium
`gpt-5.4`	OpenAI	0.8×	Standard
`gpt-5.4-mini`	OpenAI	0.25×	Standard
`gpt-5.3-codex`	OpenAI	0.6×	Standard

There is no requirement to pick an OpenAI model on this surface — you can drive a Claude model through the OpenAI wire format. Premium models require org enablement; a request for a model your org has not enabled returns a 400 invalid_request_error with hasp_code MODEL_ACCESS_DENIED. See Model Selection for cost/access details.

Non-streaming response

When stream is omitted or false, the response is a single chat.completion object:

{
  "id": "chatcmpl_01JQMSG000000000000000000",
  "object": "chat.completion",
  "created": 1775692800,
  "model": "gpt-5.4",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The HIPAA minimum necessary standard requires..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 18,
    "completion_tokens": 74,
    "total_tokens": 92
  }
}

When the model calls a tool, message.content is null and message.tool_calls carries the calls; finish_reason is tool_calls.

Streaming response (SSE)

When stream: true, the response is a text/event-stream of chat.completion.chunk objects in OpenAI’s exact format — the same shape the openai SDK’s streaming iterator expects:

data: {"id":"chatcmpl_...","object":"chat.completion.chunk","created":1775692800,"model":"gpt-5.4","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}

data: {"id":"chatcmpl_...","object":"chat.completion.chunk","created":1775692800,"model":"gpt-5.4","choices":[{"index":0,"delta":{"content":"The HIPAA"},"finish_reason":null}]}

data: {"id":"chatcmpl_...","object":"chat.completion.chunk","created":1775692800,"model":"gpt-5.4","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: {"id":"chatcmpl_...","object":"chat.completion.chunk","created":1775692800,"model":"gpt-5.4","choices":[],"usage":{"prompt_tokens":18,"completion_tokens":74,"total_tokens":92}}

data: [DONE]

The first chunk carries the assistant role, content deltas stream in the middle, a penultimate chunk sets finish_reason, a final chunk carries usage, and the stream terminates with data: [DONE]. Tool calls stream as delta.tool_calls[] entries with a stable zero-based index.

Compliance behavior

All HASP compliance checks apply — PHI Guard, BAA enforcement, and credit metering — identically to every other surface. Compliance is a property of your organization, not the endpoint shape.

PHI Guard runs on all messages per your org’s phi_mode. In redact mode PHI is replaced with placeholders before the upstream call; in block mode the request is rejected.
BAA required. No active BAA → a permission_error with hasp_code BAA_REQUIRED.
Credit metering. Credits are deducted after each successful response.

Errors

Errors are returned in OpenAI’s error envelope. HASP’s richer error vocabulary maps into OpenAI’s error type set, with the original HASP code preserved as code (lowercased) and hasp_code:

{
  "error": {
    "message": "Model 'gpt-5.5-pro' is not in the allowed model list for organization <org-id>.",
    "type": "invalid_request_error",
    "param": null,
    "code": "model_access_denied",
    "hasp_code": "MODEL_ACCESS_DENIED"
  }
}

HTTP	OpenAI `type`	Example condition (`hasp_code`)
401	`authentication_error`	Missing or revoked token (`INVALID_API_KEY`)
403	`permission_error`	No active BAA (`BAA_REQUIRED`); credits exhausted (`CREDITS_EXHAUSTED`); AI paused (`AI_PAUSED`); feature not HIPAA-eligible (`FEATURE_NOT_HIPAA_ELIGIBLE`)
400	`invalid_request_error`	Model not enabled or unknown (`MODEL_ACCESS_DENIED`); PHI blocked (`PHI_BLOCKED`); retired model (`MODEL_RETIRED`)
429	`rate_limit_error`	RPM or daily limit exceeded (`RATE_LIMITED`)
500	`api_error`	Upstream model provider error (`INFERENCE_UPSTREAM_FAILURE`)

During a stream, a mid-run failure is emitted as a terminal data: line carrying the same error envelope rather than an HTTP status change.

AI Studio

Platform

AI API

Data API

App Builder SDK

POST /v1/chat/completions

Migrating from OpenAI

Request

Body parameters

Model selection

Non-streaming response

Streaming response (SSE)

Compliance behavior

Errors

​Migrating from OpenAI

​Request

​Body parameters

​Model selection

​Non-streaming response

​Streaming response (SSE)

​Compliance behavior

​Errors

Migrating from OpenAI

Request

Body parameters

Model selection

Non-streaming response

Streaming response (SSE)

Compliance behavior

Errors