Billing and metering

Grid measures inference in input, cached-input and output tokens. Pricing publishes indicative model references in USD for new-account evaluation. The agreement for your account records the applicable serving variant, rate card, currency and effective date. Access remains by invitation.

Use the purchasing checklist to confirm eligibility, inclusions, payment terms and acceptance with your technical and procurement teams.

Published rates and eligibility#

The Models page and Pricing display the same rates. The authenticated model list determines what your account can call. A published price is a reference for an eligible serving variant; it does not enable every catalog model.

Model references contain model-specific cached-input prices. Use the explicit cached rate where listed. Where no specific cached rate is listed, cached input uses 50% of that model's input rate. The calculator counts cached input once, at the model-specific rate where one is shown.

The new-account reference is based on complete paid OpenRouter endpoint pairs reviewed on 4 October 2026. Free endpoints and temporary discounts are excluded. The comparison uses a 3:1 input/output mix and the specified reference precision where available. Serving compatibility, deployment and effective rates are confirmed before use. The pricing basis identifies the sources and scope. Proprietary LIVIA S1 Flash retains its existing Grid reference. Existing account agreements retain their effective rates. A commercial reference does not change your account agreement.

Nemotron additions need deployment and availability scoping. Nano Omni and specialist retrieval, document and voice models receive a scoped quote where no comparable paid token rate exists. A vendor release does not by itself enable that model on Grid.

Batch uses 50% of the eligible Serverless token reference. Apply that multiplier to input, cached input and output. Inference Batch and training have separate meters.

Estimating inference#

Calculate each meter separately:

uncached input / 1,000,000 × input rate + cached input / 1,000,000 × cached-input rate + output / 1,000,000 × output rate

The workload calculator lets you enter requests per month, input and output tokens per request, and the share of cached input. It includes the assumptions with an access request so the team can confirm the model, serving configuration and effective rates.

Use Copy estimate link to save the selected model, pricing mode and workload assumptions. On Pricing or Models, shortlist up to three candidates. Pricing shows a monthly token estimate for each model and can sort by Lowest workload cost, using the full input, cached-input and output mix. Models without published rates keep a Quote required label. Cost order does not establish output quality or deployment readiness.

Copy shortlist link includes the same assumptions for the selected candidates. Opening a candidate's pricing or moving between Pricing and Models keeps the shortlist and workload. Discuss these models prepares an inquiry with the candidates, access status, assumptions and indicative token totals for review before sending. Shared links recalculate estimates from the published references when opened. They do not lock a rate or grant model access. Incomplete or invalid shared assumptions show a notice and use the example workload so you can review and correct it.

Apply the mode multiplier to those token charges. Count each input token once: cached tokens use the cached rate instead of also being charged as uncached input. The agreement identifies any other billable service and tax or currency treatment.

For Smart Router, use the resolved model's rates. Your account agreement specifies any routing charge. Do not estimate a routed workload from a single model's price.

Detection models#

OpenDetector has a separate evaluation and commercial scope for AI-text, prompt injection and jailbreak checks. Its quote confirms the chargeable text unit, limits, model version, region, support and any capacity commitment. There is no published detection rate card, and inference output-token rates or Batch discounts do not price a check.

Confirm metering for long documents, repeated checks, incomplete scans and extraction before comparing total cost. See Detection commercial scope and Pricing.

Dedicated and Private#

Tokens still measure traffic, while a quote can charge GPU-hours or committed capacity. The L4 and B200 GPU-hour references are identified on Pricing. They do not establish a deployable configuration, stock or a reservation.

Dedicated qualifies the configuration, quantity, desired region, start and duration. Brazil inventory is planned for May 2027. The quote confirms configuration, region, delivery, minimum quantity and term, billable period, setup and recurring costs. It itemises storage, networking, egress, licences, support and any excess usage.

The proposal states whether model consumption is included in the commitment or billed separately. It identifies each charged service once. Private isolation and residency are defined in writing.

Projects and training#

Projects receive a scoped proposal. It separates discovery and implementation from recurring operation, included volume, model usage and any capacity commitment. The success criterion and acceptance process are part of that scope.

Training remains by invitation or a quoted engagement. Its training-token, job or capacity meter is defined for the selected path and method. Inference token rates and the Batch discount do not automatically price a training job.

Usage and reconciliation#

The Console usage dashboard breaks consumption down by model, run mode and period. Finance and operations can export the same usage as CSV with a quota ledger.

The x-grid-routed-model response header identifies the model that answered a routed call. Reconcile those usage records against the rates, inclusions and commitments in the agreement. Ask the team for the applicable quote when a needed model or service has no published rate.