$0.05184 for 8,640 AI health probes

01 / The five-cent receipt starts with explicit assumptions

A useful AI health probe should answer a narrow question: can my application authenticate, reach its configured model, and receive a usable response? It should not generate an essay every five minutes. That turns monitoring into another workload you have to fund.

Here is a planning receipt for one tiny inference request every five minutes, running continuously for a 30-day month. The token counts and rates below are illustrative assumptions, not CheapRelay prices or measured results. Replace them with your model’s current rates and the usage reported by actual requests.

• ■ Schedule: 12 requests per hour × 24 hours × 30 days = 8,640 requests.
• ■ Assumed billed input: 20 tokens per request, totaling 172,800 tokens.
• ■ Assumed billed output: 2 tokens per request, totaling 17,280 tokens.
• ■ Illustrative input rate: $0.20 per million tokens. Input subtotal: $0.03456.
• ■ Illustrative output rate: $1.00 per million tokens. Output subtotal: $0.01728.
• ■ Base inference total: $0.05184 for the month.

That number excludes the machine running the checks, logging, alerts, network charges, and any other billable token categories. It is an inference estimate, not an all-in monitoring quote.

02 / Make the probe follow your application’s real route

For an application using CheapRelay’s OpenAI-compatible interface, send the probe through the same client configuration and authentication path as production. The API base is https://api.cheapapirelay.com, with OpenAI-compatible routes under /v1. A chat-completions application would therefore test POST /v1/chat/completions.

Choose the exact model identifier your application uses. A successful call to a cheaper, unrelated model does not establish that your production model works. Keep the request minimal: one short instruction asking for a brief acknowledgment, no conversation history, no tools, and a small output budget supported by that model.

The assumed 20 input tokens are not guaranteed by a short visible prompt. Message formatting and model-specific accounting can affect billed usage. Likewise, a tiny output limit can truncate a response. Check actual usage before treating the receipt as a budget.

03 / Separate a reachable server from usable inference

Run two layers of checks. First, monitor your own application’s readiness endpoint every 30 seconds. It can verify that the process is running and configuration is loaded without making a model request. That creates 86,400 local checks in 30 days, but no inference tokens by itself. Hosting and monitoring charges can still apply.

Second, run the paid inference probe every five minutes. Require an expected success status, a parseable response, and nonempty assistant content. Do not require an exact acknowledgment unless you deliberately want to test that contract; harmless wording differences should not page anyone.

A normal non-streaming probe does not verify streaming delivery, tool calls, long contexts, or every regional network path. Add targeted checks only for capabilities your application actually needs, and budget them separately.

04 / Put retries on the receipt before enabling them

Use explicit client deadlines. For example, start with a 5-second connection timeout and a 20-second overall deadline, then adjust them to your application’s requirements. These are proposed settings, not latency benchmarks or provider guarantees.

Allow one retry after a short randomized delay for transient network errors, selected server errors, or rate limits when permitted by retry guidance. Do not repeatedly retry invalid credentials or malformed requests. A client timeout does not prove that upstream processing stopped, so retries can create additional billed work.

• ■ Budget allowance: at most 100 extra attempts per month.
• ■ At the illustrative six-millionths-of-a-dollar cost per request: $0.00060 extra.
• ■ Base receipt plus that allowance: $0.05244.

05 / Alert on failures and reconcile real usage

Record timestamp, model identifier, status, elapsed time, request ID when available, and reported token usage. Never log authorization headers. Group consecutive failures into one incident instead of sending an alert for every attempt.

CheapRelay bills per real token, including cache reads and writes, with no monthly fee or subscription. Sign up for CheapRelay, measure a small batch of probes, and replace every illustrative receipt line with observed usage and current rates. The goal is a small, predictable monitoring bill—not a tiny estimate that hides expensive assumptions.