Cost per Accepted LLM Answer: Tokens, Retries and Editorial Review

Define accepted answers before measuring anything

For a newsroom, accepted might mean verified factual statements, a complete source list and compliance with editorial style. For customer support it might mean a resolution with no escalations or policy breach. Providers do not supply a universal acceptance rate; it must be collected through consistent human or task-specific automated review. The billable quantity is the set of all requests, including attempts that do not meet the standard, subject to each platform’s actual charging policy.

Calculate each attempt in its own token categories

A Grok model can charge a distinct cached, ordinary input and output price; Mistral provides similar token separation. If a workflow makes several retries with different prompt lengths, sum their actual metered cost rather than multiplying a notional “average answer” by its first attempt. An agent may also incur web-search calls, vector database lookups, parsing and human moderation. Token cost alone is only one term of total answer economics.

A transparent accepted-response illustration

Assume 100 requests, each costing a hypothetical $0.02 in model tokens after all token categories are counted, and 80 outputs pass the defined review standard. The direct model cost per accepted response is $2/80 = $0.025. If 20 failed outputs must be retried at $0.02 and ten more then pass, the final accepted count becomes 90 with model charges of $2.40, or about $0.0267 per accepted answer. These dollar inputs are stated workload assumptions, not any provider’s measured success rate.

Human review can dominate a small token bill

Suppose editorial review consumes four minutes for each of 100 attempts. At an illustrative fully burdened labor rate entered by the operator, review labor may exceed inference cost by orders of magnitude. The ratio varies by domain and cannot be asserted from a vendor price table. Build formulas with reviewer_minutes, hourly_labor_cost, escalations, refunds and moderation overhead rather than hiding them in a model rate. Treat acceptance thresholds as stable across model comparisons.

When decisions change

A model may be worth a higher per-million-token price if it reduces wrong answers, reruns or review time—but that conclusion requires an independent evaluation on actual tasks. Store provider, exact version, locale, endpoint, prompt length, tool mode and dated price snapshot beside the production acceptance result. Do not infer guaranteed savings from vendor benchmark claims. Procurement can compare at least three workload scenarios without fabricating performance tests.

How to verify the actual invoice

Begin with a small representative workload rather than the cheapest published headline. Record the exact provider and billing channel, API model identifier, currency, account geography, plan and commitment length. The price or quota evidence in this guide is keyed to underlying measured vendor events. Do not combine proprietary creative credits, token meters or media seconds into a single currency or invent a conversion rate between them. If an enterprise or regional rate is not published, leave its price unknown and obtain a written quote.

Keep separate estimates for failed attempts, optional tools, storage, exports, human quality review and tax. Use source-verified line items from the companion CSV instead of typing permanent numbers into your forecasting formulas. Where an old plan or model was superseded, preserve its original historical record and create an additive later snapshot. This guide does not claim a product test, performance ranking, independent legal assessment or live procurement approval.

Availability, alternatives and editorial limits

Available plans, supported endpoints, trial eligibility, quota reset and contract rights may vary by region and account even when a public website displays a single global currency. Compare an alternative only with the same accepted task, input/output volume, model version and privacy requirements. Vendor-provided marketing claims are not independent evidence of quality. This historical pricing article may be cross-referenced by the model, media or enterprise procurement team, but its purpose is to explain the billing mechanics without duplicating their evaluations or inventing test scores.