/pricing · USD
Compare token rates, put your workload into the estimate, and choose how you move into production.
A clear starting point
MiniMax, DeepSeek, Qwen, GLM, NVIDIA Nemotron, Inkling, Kimi and the S Series. Separate input, cache and output rates. Serverless and eligible Batch.
Grid access is by invitation. Model rates reviewed 4–5 October 2026. Configuration and effective rates are confirmed before use.
Review the purchasing checklistCompare delivery, pricing and the requirements to prepare.
01 / Grid
02 / Dedicated
03 / Projects
Dedicated and Private requirements are scoped in writing. A capacity proposal confirms whether model consumption is included or additional. Your account agreement records effective rates.
Workload estimate
Start with a sample request. Adjust the tokens to reflect what your application sends and generates.
Examples are starting assumptions. Cached tokens are a share of total input. Output includes billable reasoning tokens where applicable.
Serverless · USD / month
$18.40
Estimated token usage
$0.17 input · $0.169 cached · $3.00 output per 1M tokens.
An indicative estimate for your evaluation. We confirm model access, serving configuration and effective rates during onboarding. Taxes, routing fees and scoped services are additional where applicable.
Your model, mode and workload go with the request.
Shared links save assumptions. The estimate uses the published references when opened; effective rates are agreed during onboarding.
Confirm the checkpoint, method and trainer for a custom job. Training receives a separate quote; this estimate covers inference.
Model pricing
USD per 1 million tokens · monthly estimates use your workload above. Indicative references reviewed 4–5 October 2026; account rates and access are confirmed during onboarding.
6 of 19 models · Highlights + your selections · Serverless · 0 of 3 shortlisted
Edit workloadMonthly estimates count the full input, cached-input and output mix. A lower estimate does not establish quality or availability. Batch uses 50% of these rates when enabled. Cached input replaces the uncached-input charge on those tokens. A 50% input rate is used when no model-specific cached reference exists. Taxes, routing and scoped services are additional where applicable.
Sources and pricing basisOpenDetector / Detection models
Scope AI-text detection, prompt injection and jailbreak checks around submitted text volume, lengths and concurrency. Detection returns signals; its usage and costs are scoped separately from generated output tokens.
Explore the detection modelsThe proposal confirms the model version, chargeable unit, text limits, throughput, region and support. It also states how long documents, repeated checks, incomplete scans and private capacity are charged.
Evaluation by request. There is no published detection rate card or assumed Batch discount. Deployment and effective rates are confirmed before access.
Prepare the cost comparisonAdapting a model?
Choose custom training, your own loop or managed jobs. Confirm dataset readiness, checkpoint eligibility, evaluation and the serving plan before pricing the job.
Evaluate before you scale
Choose a representative workload. Compare quality, response time and cost per accepted result. Set the acceptance criteria before you increase volume.
01 / Baseline
Bring sample inputs, expected outputs and monthly volume.
02 / Compare
Measure candidate models against your quality and cost targets.
03 / Commit
Agree on the configuration, rates, support and rollout scope.
Dedicated · committed capacity
For teams that need a defined configuration, quantity, region and duration. The proposal specifies the resources, service, delivery and total commitment.
Prepare a capacity briefDedicated inventory: Brazil, planned for May 2027. Configuration and quantities are confirmed in the quote.
Rate card v1 · September 2026 · USD. These are metering references for eligible GPU pools. A Dedicated quote confirms the GPU configuration, quantity, region, minimum term, billing basis and inclusions. A listed rate does not reserve capacity or confirm a delivery date.
Projects · a production outcome
For teams that need implementation, integration and operation. We scope the problem, workload volume, systems and success criterion, then separate delivery costs from recurring usage.
Prepare a project briefScope retrieval, document processing or voice with NVIDIA's specialist models. Model and deployment availability are confirmed in the proposal.
These services have workload-specific units and quotes. Free model previews do not establish a commercial rate.
Paid endpoint references, compared at a 3:1 input-to-output mix. Your workload and serving configuration determine the comparison; effective rates are confirmed before use.
We select a complete input and output rate pair from the lowest-cost active paid OpenRouter endpoint at a reference mix of three input tokens to one output token. We use standard list rates, excluding free endpoints and temporary discounts. We compare at the stated reference precision where specified. Grid serving precision and eligibility are confirmed for your deployment. A different token mix, precision or deployment can change the comparison.
We retain the endpoint's cached-input rate when published. Otherwise the Grid reference uses 50% of input. LIVIA S1 Flash retains its own Grid rate; other S Series generations receive their agreed rates. Nemotron Nano Omni and Nano 4B have no comparable paid listing in this review and receive scoped quotes. Context and feature limits are confirmed for your deployment.
These are indicative new-account references in USD. Your agreement records the model, serving configuration, effective rate and date, billing currency and validity. Existing agreements retain their agreed rates. Tax and currency treatment are confirmed in writing.
Grid access is by invitation. The authenticated model list determines what your account can call. Dedicated and Private proposals confirm the deployment region, residency and delivery window.
Capacity and project proposals state minimum quantities and terms, setup fees, resources and support. Storage, networking, egress, licences and excess usage are itemised where applicable.
Input, cached input and output are separate token meters. A capacity quote explains whether model consumption is included or additional. Routing, training and other scoped services have their own agreed rates.