Chat Completions

POST /v1/chat/completions in the OpenAI format: what is supported and what is ignored.

POST https://api.egrtgghfghtytgb.space/v1/chat/completions takes the OpenAI Chat Completions format for Claude models. Unspent translates the request to the Messages API and the answer back.

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://api.egrtgghfghtytgb.space/v1", apiKey: process.env.UNSPENT_API_KEY });

Prompt caching, extended thinking, web search and client tools have no equivalent in this format. Use the Messages API for them.

Supported

OpenAI	Becomes
`system` and `developer` messages	Joined with a newline into `system`
`user` and `assistant` text	Text blocks
`image_url` parts: data URLs and https URLs; JPEG, PNG, GIF, WebP	Image blocks
`tools` of type `function`, `tool_choice`, `parallel_tool_calls`	Tools with `input_schema`, `tool_choice`
`assistant.tool_calls` and `tool` messages	`tool_use` and `tool_result` blocks
`max_completion_tokens` or `max_tokens`	`max_tokens`, 4,096 if neither is set
`stop`	`stop_sequences`
`n`	Only `1`
`stream`, `stream_options.include_usage`	`chat.completion.chunk` events, then `data: [DONE]`

A request holds its worst-case cost before it runs, and max_tokens is part of it. Without max_tokens a request holds 4,096 tokens of output.

Ignored

These fields are accepted and dropped, so existing code keeps working:

temperature, top_p, response_format, logprobs, top_logprobs, seed, reasoning_effort, frequency_penalty, presence_penalty, logit_bias, user, metadata, store, service_tier, prediction, modalities, audio.

temperature and top_p are dropped because current Claude models accept only their default sampling settings. Unknown fields are ignored as well.

Rejected

functions, function_call and messages with role function return 400 unsupported_parameter. That is the legacy function calling: dropping it silently would mean the model never calls your functions. Use tools and tool_choice.

Response

choices[0].message holds the text and any tool_calls.
finish_reason is stop for a normal end or a stop sequence, length when max_tokens is reached, tool_calls when the model calls a tool and content_filter for a refusal.
usage.prompt_tokens counts input plus cache reads and writes, completion_tokens counts output with thinking, prompt_tokens_details.cached_tokens counts cache reads.

Every response, and the last chunk of a stream, carries an unspent field with the request's cost and your balance after it:

json

"unspent": {
  "request_id": "3f9c0e…",
  "credits_charged": "0.01035",
  "available": "24.98965"
}

Errors

Errors use the OpenAI format:

json

{ "error": { "message": "Not enough credits. …", "type": "insufficient_quota", "code": "insufficient_credits" } }

The full list: Errors.