Limits and errors

Limits

Limit	Value
Tokens per request	131,072, state and questions together
Questions per request	64
Options per choice	Tested up to 151
Images per request	8, as base64 data URLs
thinking_budget	16 to 4096 tokens
Span max_tokens	1 to 96
Spans max_items	1 to 64

Errors

Status	When	Body
400	model is missing	{"error": "Model name not provided"}
400	Unknown model name, such as jev-latest	{"error": "Model not found"}
400 / 422	Invalid request: unknown thinking value, bad question shape, state too long, raw text state on chat completions	{"error": {"code": ..., "message": "..."}}
401	No Authorization header, or not in the form Bearer sk-...	{"error": {"code": "authentication_error", "message": "Authorization header missing or invalid format. ..."}}
403	API key not valid	{"error": {"code": "authentication_error", "message": "Access forbidden"}}
429	Rate limit for your key	No Retry-After header. Wait a second or two and retry.

Reading validation errors

The message of a validation error is itself a JSON string. Parse it to get the reason, for example thinking: one of "off", "auto", "auto_forced", "always".

import json

resp = requests.post(url, headers=headers, json=body, timeout=30)
if resp.status_code in (400, 422):
    error = resp.json()["error"]
    if isinstance(error, dict):
        try:
            reason = json.loads(error["message"])
        except (TypeError, ValueError):
            reason = error["message"]
        print("Invalid request:", reason)

Retrying

Retry 429 and 5xx responses with backoff. A 429 has no Retry-After header, so wait a second or two before the first retry and longer after each one. Don't retry 400, 401, 403 or 422: fix the request or the key first.

Getting help

If you think a response is wrong or the service is misbehaving, contact ASI:Cloud support.