Chat completions

POST https://heyaskr.ai/v1/chat/completions

Create a model response for a conversation. The request is the OpenAI shape. This page is precise about which fields we read, because the rest are dropped without a warning.

Request fields we read

Everything else is dropped silently: temperature, top_p, tools, tool_choice, response_format, stop, n, seed, plugins. A request that sends them succeeds; they just have no effect. Tool calling and structured output are on the roadmap. Web search is not a tool you define: it is the web_search field above.

Messages

• Roles other than system, user and assistant return 400 Unsupported message role.
• content is a string, or an array of parts: text, image_url, file, input_audio and video_url, each inline or by file_id. Any other part type returns 400 Unsupported content part. A message that is only parts is sent as it is. Pictures and files.
• Only the last 40 messages are sent to the model. Trim on your side if you want control over what is kept.
• A conversation may fill the model's own context window, less room for the reply: about 1 million tokens on the largest models, such as Kimi K3, at most 3.2 million characters and the last 400 messages. A model with a small window keeps 120,000 characters and 40 messages. Over the limit returns 400 The conversation is too long.
• No usable messages returns 400 No usable messages were provided.

Example

curl https://heyaskr.ai/v1/chat/completions \
  -H "Authorization: Bearer $ASKR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 300,
    "messages": [
      {"role": "system", "content": "Answer in one sentence."},
      {"role": "user", "content": "Why is the sky blue?"}
    ]
  }'

Response

{
  "id": "chatcmpl-9c2b...",
  "object": "chat.completion",
  "created": 1789516800,
  "model": "claude-sonnet-5",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Sunlight scatters off air molecules, and blue scatters most." },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 21,
    "completion_tokens": 14,
    "total_tokens": 35,
    "completion_tokens_details": { "reasoning_tokens": 0 }
  },
  "askr": { "credits_charged": 0.192 }
}

How the cost is worked out

Before the call, credits are reserved for the worst case: your estimated input tokens at the model's input rate, plus max_tokens at its output rate. When the model finishes, the reservation is settled to the real cost, using the gateway's own figure for the call. Nothing else is added: there is no platform fee in the first version. Holds and settlement.

Pictures and files in a message

A user message's content can be an array of parts, in OpenAI's shape. Five kinds are read. A part carries its bytes inline, or names an upload by file_id from POST /v1/files; an upload is the way to send anything big, or anything more than once.

The request body on this endpoint is 12 MB, so an inline part over about 9 MB does not fit: upload it. An inline part goes to the model as it is, so it works only on a model whose inputs on GET /v1/models include that kind. A part by file_id is routed per model: the file itself where the model takes the kind, its words where it does not, so a PDF by file_id works on every model. Files has the rules and the costs.

{
  "role": "user",
  "content": [
    { "type": "text", "text": "What do these say, and does the clip match?" },
    { "type": "file", "file": { "file_id": "3f2a9c1e-7b4d-4e0a-9c1f-2d8e6a5b4c3d" } },
    { "type": "file", "file": { "filename": "notes.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjcK..." } },
    { "type": "input_audio", "input_audio": { "data": "UklGRiQAAABXQVZF...", "format": "wav" } },
    { "type": "video_url", "video_url": { "url": "data:video/mp4;base64,AAAAIGZ0eXBpc29t..." } }
  ]
}

Cost. A file is input tokens at the model's input rate. The hold reserves 1,600 tokens a picture, 1,600 a PDF page, 32 a second of sound, 100 a second of video, and the text's own count where a file goes as words; it settles to the gateway's real figure. The three newest user messages that name uploads carry them in full; older ones get a one-line note in the file's place, so a PDF at the top of a long conversation is not paid for on every call.

Refused before anything runs. Nothing is reserved for these. Errors.

Limits

Requests

120 per minute per IP address

Body size

12 MB on this endpoint, for inline pictures and files; uploads go to /v1/files at up to 100 MB

Output

max_tokens up to 8,192; default 4,096

Deadline

120 seconds upstream. Past it you get 504 and the reservation is charged, because the model did run. Errors

Search the web

Send "web_search": true and the model can search the web and read pages during the call, exactly as the globe in the workspace does. The model is given two tools of ours, web_search and fetch_page, decides for itself whether to use them, and answers with numbered [n] marks. Per call: at most 3 searches, 2 pages read and 4 model calls, the last of them made without tools. It works on a model whose entry on GET /v1/models says supportsTools: true, streamed or not. Your own tools are still dropped.

curl https://heyaskr.ai/v1/chat/completions \
  -H "Authorization: Bearer $ASKR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "web_search": true,
    "messages": [{"role": "user", "content": "When did askr add web search?"}]
  }'

{
  "id": "chatcmpl-9c2b...",
  "object": "chat.completion",
  "created": 1789516800,
  "model": "gpt-5.6-sol",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Search the web went live on 23 September [1].",
        "annotations": [
          {
            "type": "url_citation",
            "url_citation": { "url": "https://heyaskr.ai/changelog", "title": "Changelog", "start_index": 41, "end_index": 44 }
          }
        ]
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 3412,
    "completion_tokens": 58,
    "total_tokens": 3470,
    "searches": 1,
    "fetches": 0,
    "search_credits": 5
  },
  "askr": { "credits_charged": 5.31 }
}

Streamed, the chunks are the ones on Streaming, with three differences: content from every model call arrives in order with a blank line between calls; a : askr comment line appears while a search or a page read runs, so an idle proxy does not close the stream; and the last data frame carries delta.annotations, finish_reason, usage and askr together.

data: {"id":"chatcmpl-9c2b...","object":"chat.completion.chunk","created":1789516800,"model":"gpt-5.6-sol","choices":[{"index":0,"delta":{"role":"assistant","content":""},"finish_reason":null}]}
: askr
data: {"id":"chatcmpl-9c2b...","object":"chat.completion.chunk","created":1789516800,"model":"gpt-5.6-sol","choices":[{"index":0,"delta":{"content":"Search the web went live"},"finish_reason":null}]}
...
data: {"id":"chatcmpl-9c2b...","object":"chat.completion.chunk","created":1789516800,"model":"gpt-5.6-sol","choices":[{"index":0,"delta":{"content":"","annotations":[{"type":"url_citation","url_citation":{"url":"https://heyaskr.ai/changelog","title":"Changelog","start_index":41,"end_index":44}}]},"finish_reason":"stop"}],"usage":{"prompt_tokens":3412,"completion_tokens":58,"total_tokens":3470,"searches":1,"fetches":0,"search_credits":5},"askr":{"credits_charged":5.31}}
data: [DONE]

Cost. Each search is 5 credits. The text the model reads is billed as input tokens at the model's rate, like the rest of the conversation, and every extra model call sends the conversation again. The hold is raised by 15 credits plus the most text the searches, the pages and the extra calls could add at the model's input rate; a 402 names that figure. It settles once, to the searches that ran plus the real token usage across every call. A call cut short is settled on what streamed, at most the hold. A call that ends with no content releases the hold and charges nothing, searches included. Each account may start 6 searching calls a minute, workspace and API together.

Refused before anything runs. Nothing is reserved for these. All three carry "param": "web_search" and the workspace's code in code, so one code means one thing on both surfaces. Errors.