Errors & reliability

Status codes, and the one behaviour worth handling in code: a cold start.

Status codes

Error responses are JSON with an error message, for example:

{ "error": "Invalid or revoked API key." }

Cold starts & warm‑up

To keep costs low, idle compute is powered down. The first request after an idle period wakes a node and loads the model, which can take up to a minute or two. During that window the API returns 503 with "code": "no_node" instead of a wrong or fake answer.

This isn't an error to surface to your users — it just means “not ready yet, try again shortly.” Retry the same request a few times with a short delay; a warm node answers in a couple of seconds.

async function chat(message, key) {
  for (let attempt = 0; attempt < 6; attempt++) {
    const res = await fetch("https://stevai.com/api/v1/chat", {
      method: "POST",
      headers: { "Authorization": `Bearer ${key}`, "Content-Type": "application/json" },
      body: JSON.stringify({ message }),
    });
    const data = await res.json();
    if (res.ok) return data.reply;
    if (res.status === 503) {            // warming up — wait and retry
      await new Promise((r) => setTimeout(r, 5000));
      continue;
    }
    throw new Error(data.error || "Request failed");   // 400 / 401 — a real error
  }
  throw new Error("The network did not warm up in time. Try again.");
}

Idempotency & retries

• Retrying a 503 is safe — a request that didn't produce an answer isn't billed.
• Back off a few seconds between attempts; the model keeps loading between your retries, so a later attempt lands.
• Don't retry a 400 or 401 — those are real problems (bad input or bad key), not transient.

Cost & billing

Every successful response reports the real cost of that request in its usageobject, and that cost is metered to the API key's account. Only successful requests are billed.