Agent/MCP Audit Sprint

Provider bill to product display check

OpenRouter Actual Cost Reconciliation

Compare what your product showed a user with the OpenRouter provider actual cost from usage.cost or generation total_cost. The goal is to separate display bugs from real spend leaks: generation id idempotency, per-response versus cumulative totals, cache-read pricing, reasoning tokens, retry dedupe, and multi-call step aggregation.

Run reconciliation Use calculator Balance guardrail 402 brownout TokenMeter agent review Request OpenRouter review
Free toolProvider actual cost vs product display
QuickUSD $99 OpenRouter Cost Sanity Check
FocusedUSD $299 OpenRouter Agent Cost-Control Review
EmergencyUSD $1,000 AI Cost Spike Emergency Sprint
Start RulePayment only after written scope acceptance

Reconciliation packet

Turn a cost mismatch into scoped evidence

Use only sanitized values: product-displayed cost, OpenRouter actual cost, generation count, token totals, cache-read tokens, reasoning tokens, and retry or replay counts. Do not paste private prompts, API keys, customer data, account ids, raw production traces, or billing screenshots with sensitive values.

Ratio24.0x OpenRouter actual vs product display
Delta+$0.20 provider actual above product display
FitUSD $299 OpenRouter Agent Cost-Control Review
GuardrailPayment only after written scope acceptance; this packet is not a payment request.
Open GitHub intake

What gets reconciled

The paid review treats the provider actual-cost field as a ledger input, then checks whether the product display, budget gate, and audit trail are using the same scope.

View $299 focused review View $1,000 emergency sprint
Provider actual: reconcile usage.cost, generation total_cost, model, provider, and token count fields.
Generation id: make every generation id idempotent so refreshes, polling, retries, and replays do not double count.
Scope: keep per-response, per-step, per-run, and cumulative totals separate in code and UI labels.
Aggregation: preserve multi-call step totals without losing the underlying generation records.
Fallback pricing: handle cache-read tokens, reasoning tokens, unknown price rows, and provider fallback without inventing false certainty.
Retry dedupe: charge failed, retried, streamed, or replayed calls according to the provider record, not the UI event count.

Routing

Use the mismatch shape to choose scope

A tiny mismatch can be a $99 sanity check. A repeatable product-versus-provider gap usually belongs in the $299 focused review. A live bill spike or large absolute delta routes to emergency containment.

USD $99: one display label, pricing-unit, or cost-field sanity check before launch.
USD $299: one OpenRouter-backed workflow with actual-vs-estimated fields, generation id dedupe, cache/reasoning handling, budget-gate alignment, and audit trail fixes.
USD $1,000: active production spike, large unexplained provider charge, or urgent containment before more traffic arrives.
Evidence boundary: share redacted cost fields, token totals, generation ids if safe, and summaries only; never share prompts, keys, customer data, or private traces in a public issue.