Tokens: Units of text or other content processed by a model. Tokenization depends on the model and input type. Billing is typically reported in input tokens, output tokens, and — where applicable — cached input tokens.
Related: input tokens, output tokens, cached tokens.
Explore related resources
Prompt caching: Reuse of eligible previously processed prompt content, typically at a reduced token rate. Provider eligibility rules, cache retention windows, and billing categories vary. Anthropic and OpenAI both offer prompt caching; the mechanism and rates differ.
Check the provider's current caching documentation before relying on cached rates.
Explore related resources
Cost attribution: Assigning AI usage costs to the projects, teams, environments, or operations responsible for them. Model Meter attributes costs by setting projectId in the SDK init() call and by labeling individual calls with an operation field.
Related: project, operation, cost per request.
Explore related resources
Latency: The time a request takes to return a response. Time to first token (TTFT) and total response time measure different stages. Model Meter records latency in milliseconds per request.
Compare latency alongside quality and cost when evaluating model switches.
Explore related resources
Cost per request: Total measured cost divided by the number of requests within a defined workload and time period. Use consistent billing categories (same model, same token type) when comparing across models or time periods.
Related: cost attribution, effective-dated pricing.
Explore related resources
Estimated cost: A cost value calculated by Model Meter using a pricing catalogue record and the token usage reported in the request. Clearly distinguished from provider-reported cost. Estimates may differ from invoices due to pricing changes, discounts, credits, or batch rates not captured at the request level.
Related: provider-reported cost, reconciled cost.
Explore related resources
Provider-reported cost: Cost figures returned by the provider's usage or billing API and imported into Model Meter. Always stored separately from request-level estimates. The billing reconciliation feature compares the two figures side-by-side without mixing them.
Related: estimated cost, billing reconciliation.
Explore related resources
Effective-dated pricing: A pricing catalogue where each price record has an effective start date. When a provider changes model prices, a new record is added with the new effective date. Historical cost records retain the pricing version that was active at the time of the request, so past estimates are not silently rewritten.
Related: estimated cost, cost provenance.
Explore related resources
Idempotent ingestion: An ingestion system where retrying the same event — due to network errors, timeouts, or SDK flush retries — does not inflate usage counts or cost totals. Model Meter uses a sourceEventId per event and database unique constraints to deduplicate retried submissions.
Related: SDK, ingestion API.
Explore related resources
Cost-to-quality: A comparison of model cost per request (or per successful task) against a task-specific quality metric, evaluated on the same dataset. Model Meter's cost-to-quality feature runs candidate models on an evaluation dataset and recommends lower-cost alternatives only when quality and latency constraints are met.
Related: model evaluation, cost per request, latency.
Explore related resources
OpenTelemetry (OTel): An open standard for instrumenting applications with traces, metrics, and logs. The Vercel AI SDK emits gen_ai.* OTel spans for every LLM call. Model Meter's SDK includes an OTel span exporter that captures these spans and sends them to the Model Meter API asynchronously — without sitting in the critical path of the AI call.
Related: gen_ai.* spans, SDK, Vercel AI SDK.
Explore related resources
Billing reconciliation: The process of comparing Model Meter's request-level estimated costs against provider-issued invoice data. Reconciliation records are always stored separately from estimates. A reconciliation report shows both figures side-by-side and flags differences due to price changes, credits, adjustments, or missing pricing records.
Related: estimated cost, provider-reported cost.
Explore related resources
Tenant isolation: The guarantee that one workspace cannot access another workspace's data. Model Meter enforces tenant isolation at the application layer (every query is workspace-scoped) and at the database layer via PostgreSQL Row-Level Security (RLS) policies.
Related: workspace, PostgreSQL RLS.
Explore related resources

A shared language for AI costs.

Lock in the best launch pricing.