Questions people ask when an AI bill looks wrong.
Answers about reliable and accountable LLM inference, BYOK versus Managed, eligible failed-call credits, and the receipts that prove what happened.
What is Inferock?
Inferock is reliable and accountable LLM inference. You route selected model calls through Inferock; it records independent per-call receipts and, in Managed mode, can apply bounded service credits for eligible objective provider failures under agreed terms.
How are Bring Your Own Keys and Managed different?
Bring Your Own Keys shows failed calls and cost on your own provider bill. Managed can apply bounded service credits for eligible objective failures because Inferock operates the provider relationship as part of the service.
What happens when an inference call fails?
Inferock records the request path, provider response, timing, usage, charge context, and failure evidence. The report shows what failed, what it cost, and whether the row is report-only or eligible for Managed service credit.
What counts as an objective failure?
Downtime, provider timeouts, empty output, truncated output, invalid structured output or tool call arguments when a schema exists, and material billing or cache anomalies with observable evidence.
What is reported but not automatically credited?
Latency without a configured threshold, model drift without an agreed baseline, refusals that may be appropriate, factuality without ground truth, and policy or security signals that need review.
Billing evidence
How do I audit my AI or LLM bill?
Run inferock-bench locally, send metered provider API traffic through it, and review the receipt. The receipt leads with spent dollars, bill-bounded money loss, time loss, and invoice-check exposure, then keeps provider-recognized recovery, recognition gap, and invoice-check exposure detail separate. It preserves token usage, cost, failure, retry, and billing-integrity evidence; it does not replace provider invoice records.
Why are my OpenAI or Claude API costs so high?
Start by separating usage volume, token mix, retries, failed calls, cache behavior, and pricing. `inferock-bench` can record metered traffic that passes through the local proxy and show per-call token usage, cost, failures, retries, cache signals, the spent/money-loss/time-loss/invoice-check-exposure headline, and the invoice-check exposure detail. It cannot explain provider charges for traffic it did not see, and it does not replace the provider invoice.
Can Inferock explain charges when my API usage dashboard shows zero, or when I was not using the API?
Only for traffic it observed. If calls passed through `inferock-bench`, the local receipt can show the measured request, response, usage, timing, retry, and cost evidence. If a provider key was used outside the proxy, the bench cannot reconstruct that unobserved traffic. Revoke exposed keys, preserve provider dashboard and invoice records, and use `inferock-bench` to measure future metered calls.
Was I billed for a failed, timed-out, or 504 API call?
Inferock records the provider response, usage fields, timing, delivered content, error status, and price evidence for the call. When charge evidence is present, it can show bill-bounded money loss for a failed, incomplete, timed-out, or 504 call. When the evidence is only an invoice-check signal, such as cache discount at risk, the receipt labels it as invoice-check exposure instead of money loss. If the provider exposes too little charge evidence, the receipt labels that limit instead of inventing a refund claim.
Can Inferock prove OpenAI overcharged tokens?
Inferock can preserve local evidence for OpenAI token, usage, retry, and failure questions. It reports evidence strength, separates bill-bounded money loss from provider-recognized recovery and recognition gap, and keeps invoice-check exposure out of the money-loss total instead of declaring every mismatch an overcharge.
Can Inferock detect an Anthropic billing error?
Inferock can record Anthropic usage fields, delivered output, pricing evidence, and token cross-check signals where available. The result is a receipt-grade evidence trail, not an automatic accusation.
How do I measure Claude or GPT token usage locally?
Use inferock-bench as a localhost proxy and point your SDK base URL at it. Provider keys stay local, and the bench records token usage and cost evidence for supported OpenAI, Anthropic, Gemini, and pinned OpenRouter calls.
Is inferock-bench an LLM cost tracking proxy?
Yes. inferock-bench is a local LLM cost-tracking proxy for development and evaluation traffic. It turns provider calls into per-call receipts with spent dollars, bill-bounded money loss, time loss, invoice-check exposure, failure evidence, and billing-integrity signals.
Can inferock-bench stop a runaway agent or retry loop from burning API credits?
It is a measurement proxy, not a provider-wide spending firewall. For traffic routed through the local bench, it records retries, duplicate request evidence, token usage, timing, and cost so a retry loop or agent run leaves a receipt. The built-in test flow shows an estimate and spend cap before it starts provider calls. Your provider account still needs its own key controls, quotas, and billing limits.
Is this an AI provider reliability benchmark?
The public benchmark measures observable reliability and billing signals on real provider traffic. It is not a provider ranking unless a run explicitly reports comparable measured data and coverage.
Can Inferock show why my OpenAI or Claude API response is slow?
It records timing for calls routed through the bench and can report latency against configured thresholds. It can show that a call was slow, timed out, or failed from the client-visible evidence. It cannot see provider internal queues or prove an internal cause unless the provider exposes that evidence.
Can Inferock tell whether a model got worse?
It can record the requested model, exposed model/version fields, output shape, refusal signals, latency, and configured drift probes. A quality change becomes useful evidence only when there is an approved baseline, threshold, or ground truth. Broad complaints like "the model got worse" stay report-only until the receipt has that baseline.
Does Inferock measure hallucinations?
Not as a general in path failure class. Factuality needs ground truth: a reference answer, retrieval corpus, database result, human label, or downstream verifier. Without that, Inferock can only flag weaker signals.
What does the loss report include?
Failure type, event count, estimated cost, evidence, and remedy status. The Accountability page includes a sanitized sample scorecard.