Our scrubbing & audit layer, running Sigmix-authored firmware — with an optional PCIe card for the appliance.
Explore Sigmix Guard
Every request served on-box, sealed from the internet. KSA Cloud runs elastic and in-Kingdom.
See your postureEverything here applies to both routes. Each route's own page has its models, prices and samples.
https://inference.sigmix.sa/ksa-sovereign-route-only/v1Send the key as Authorization: Bearer, followed by the key. x-api-key is accepted too, for SDKs that send it.
Sovereign keys start sgx-ksa-, international keys sgx-intl-. A key on the other route's path is refused with wrong_route.
We keep a fingerprint of each key, never the key. Revocation takes effect on the next request.
Authorization: Bearer sgx-ksa-… # sovereign key Authorization: Bearer sgx-intl-… # international key
Request and response shapes follow OpenAI's, so the official SDKs work by changing base_url and api_key.
chat/completions, embeddings, audio/speech, audio/transcriptions, audio/translations, images/generations, and GET models.
x-request-id on every response; x-sigmix-route; x-sigmix-charge-usd with the exact charge on non-streamed responses (a stream's charge is on the Usage page); x-sigmix-max-tokens-clamped when max_tokens was above the model's limit.
Use the id from the route's model list or GET <base>/models. The same id can exist on both routes as two separate models with two prices.
The Responses and Assistants APIs, file uploads and fine-tuning. Image URLs in chat work where the model fetches them; on some international models only base64 data URLs are accepted.
temperature, top_p, n, stop, seed, frequency_penalty, presence_penalty, logit_bias, logprobs, top_logprobs, response_format, tools (function tools only), tool_choice, parallel_tool_calls and stream_options; reasoning_effort, reasoning_format and include_reasoning on models that reason. user, metadata, store and service_tier are accepted and not passed on. Any other parameter, or one the model cannot honour, is refused with invalid_request naming it — never silently dropped. seed is best effort. Tool-call ids are issued by us, in the call_ form.
stream: true on chat completions returns server-sent events in OpenAI's chunk format, ending with data: [DONE].
stream work at any balance that covers them.max_tokens — is reserved before the first token. max_tokens is always sent to the model, so the ceiling holds.stream_options.include_usage to receive the usage chunk; billing uses it either way.Errors use OpenAI's shape, with a code that says whose problem it is. Nothing is charged for a request that failed.
{
"error": {
"message": "Your available balance does not cover this request…",
"type": "insufficient_quota",
"code": "insufficient_credit",
"param": null,
"request_id": "6f0c…"
}
}| HTTP | Code | Whose | Meaning | What to do |
|---|---|---|---|---|
401 | invalid_api_key | Yours | The key is missing, malformed, unknown or revoked. | Send Authorization: Bearer with a current key from the console. |
403 | wrong_route | Yours | The key belongs to the other route. | Use the key on its own base path, or create a key for this route. |
403 | account_suspended | Yours | The account is suspended or closed. | |
400 | invalid_request | Yours | The body is malformed or a field is invalid. | Read error.param and error.message. |
404 | unknown_endpoint | Yours | There is no endpoint at that path. | Check the path against the endpoints table on the route page. |
404 | model_not_found | Yours | That model is not published on this route. | GET <base>/models lists what is. |
403 | model_not_allowed | Yours | The key is restricted to other models. | Edit the key's allowed models. |
400 | modality_mismatch | Yours | The model is called on a different endpoint. | Use the endpoint named in the message. |
413 | payload_too_large | Yours | The body is over 16 MB, or the audio over 25 MB. | Send less, or split the audio. |
400 | context_length_exceeded | Yours | The messages are far longer than the model's context window could hold. Refused before anything is held or sent; not charged. | Shorten the conversation, or use a model with a larger context window. |
402 | insufficient_credit | Yours | Your available balance does not cover the most this request could cost. | Add credit, or lower max_tokens. |
402 | paid_credit_required | Yours | Free-trial credit cannot pay for this model, and your paid balance does not cover the most this request could cost. | Top up, or use a model included in the free trial. |
403 | streaming_not_allowed | Yours | Streaming is off for this key or account, or the balance is under the floor. | Send without stream: true, or add credit. |
429 | rate_limited | Yours | Too many requests this minute for this key. | Wait for Retry-After seconds. |
429 | concurrency_limit | Yours | Too many requests at once for this key, or more than four requests over 1 MB still uploading at once from your account. | Wait for one to finish, or raise the key's limit. |
429 | spend_cap_reached | Yours | A daily, monthly or per-request cap you set has been reached. | Raise the cap on the key, or wait for the window. |
503 | gateway_disabled | Paused | The operator has paused the API. | Retry after Retry-After seconds; see the status page. |
503 | route_disabled | Paused | The operator has paused this route. | Retry later. A sovereign request is not moved to the other route. |
503 | modality_disabled | Paused | The operator has paused this endpoint. | Retry later. |
503 | model_disabled | Paused | The operator has paused this model. | Retry later, or use another model. |
503 | capacity_exhausted | Ours | Every endpoint for this model is busy, or the API is receiving too many large requests at once. | Retry after a moment. |
503 | upstream_unavailable | Ours | No endpoint on this route can serve the model right now. | Retry; see the status page. |
504 | upstream_timeout | Ours | The model did not answer in time. | Retry, or lower max_tokens. |
502 | upstream_error | Ours | The model endpoint failed. | Retry. Not charged. |
400 | upstream_rejected | Yours | The model refused the request — too long, or a value it does not accept. The reason is given in our words, with the parameter when there is one. | Read the message; shorten or adjust the request. |
504 | deadline_exceeded | Yours | The request reached the 15-minute limit. A streamed answer stops there and only what was generated is charged; a request that had not started answering is not charged. | Lower max_tokens, or split the work. |
503 | service_restarting | Ours | The API was restarting. A streamed answer in progress stops and only what was generated is charged; a request that had not started is not charged. | Retry after Retry-After. |
500 | internal_error | Ours | Something failed on our side. | Retry, and quote request_id if it persists. Not charged. |
Requests per minute per key: the platform default unless you set your own on the key. Exceeding it returns 429 with Retry-After.
JSON bodies up to 16 MB; audio up to 25 MB; up to 8 choices (n) per chat request and 10 images per image request.
If an endpoint fails before answering, the gateway retries on another endpoint on the same route. Retry yourself on 429, 503 and 504; never on 4xx.
For each request: when, which key, which model and route, the token, character, second or image counts, the charge, the latency and the outcome. Never the prompt, the completion, the audio or the image.
On the international route, the service that answers a request processes it under its own terms.
Create an account, make a key for the route you need, and paste a sample from its page.