Documentation

Latency and currency notes for Latin American teams in the NexoRouter documentation.

Latency and currency notes for Latin American teams

NexoRouter bills prepaid balance in USD. Latency depends on your backend region, the selected model, prompt size, and upstream conditions. This page does not publish latency benchmarks or unpublished prices.

Currency and prepaid balance

  • Public credit is prepaid in USD. See Pricing and Pricing and billing.
  • Public docs use 1 USD = 500000 quota.
  • Public credit purchases start at $5. Checkout lists API credit and the 5.5% service fee (minimum $0.80) separately.
  • Your card issuer, bank, or local payment method may add its own FX, tax, or processing amount. NexoRouter does not convert model rates into local currency on the catalog page.
  • Billing is the source of truth for currently available payment methods.

Estimate test spend from the model profile's published unit price, then confirm the actual charge in Usage Logs.

Measure latency from your region

Do not copy a latency number from a blog, a vendor page, or this guide.

  1. Run the backend (or Replit, or a small VPS) in the region closest to your users when you can.
  2. Send the same small Chat Completions request to the exact model ID you plan to use.
  3. Record end-to-end request time in your client and the latency field in Usage Logs.
  4. Repeat at the hours you actually serve customers.
  5. Compare a second catalog ID with the same prompt before moving infrastructure.

If one model is slow, switch models or shorten the prompt before assuming the gateway region is the only cause. See Which server region should I use?.

Procurement checklist

  • One production API key with a budget and model scope.
  • A documented catalog ID, not a display name.
  • Prepaid balance enough for the first test window.
  • Usage Logs access for cost, latency, status, and request ID.
  • A fallback model ID that already passed the same test set.
  • Timeouts high enough for slow models; retries only for transient errors.