AWS Comprehend cost: per unit of text analyzed

Quick answer

Comprehend bills per unit of text (100 characters, with a per-request minimum of 3 units, so 300 characters) for each NLP operation, sentiment, entities, key phrases, language, with higher rates for custom classification and entity models (which also charge for training and endpoint hours). Cost scales with text volume times the number of analyses, so analyzing only what you need and batching short texts are the levers.

Amazon Comprehend performs natural-language processing, sentiment, entity extraction, key phrases, language detection, and more. It is priced per unit of text, where a unit is 100 characters, with a minimum charge per request. The cost scales with how much text you analyze and how many operations you run on it.

Per text unit, per operation

Each operation bills per 100-character unit, with a minimum of 3 units (300 characters) per request. So very short texts still cost the minimum, and running several operations (sentiment plus entities plus key phrases) on the same text bills for each. Rates fall at volume, and a free tier covers initial usage.

Custom models cost more

Type	Cost
Built-in operations	Per text unit
Custom classification / entity models	Training + per-unit inference + endpoint hours

Custom classification and custom entity models add training cost and, for real-time inference, an endpoint billed per hour while provisioned, the same idle-endpoint trap as other ML endpoints. Use asynchronous (batch) inference where real-time is not needed to avoid the endpoint hours.

Controlling Comprehend cost

Analyze only the text and operations you need, batch short texts to avoid paying the per-request minimum repeatedly, use asynchronous inference for custom models where real-time is not required (no always-on endpoint), and cache results. Because the per-request minimum penalizes many tiny requests, batching is especially effective, the send-less-and-batch discipline common to per-unit services.

FAQ

How is AWS Comprehend priced?

Per unit of text (100 characters) for each NLP operation, with a per-request minimum of 3 units (300 characters), so very short texts still cost the minimum. Running several operations on the same text bills for each. Custom models add training, per-unit inference, and endpoint-hour charges. Rates fall at volume.

How do I reduce Comprehend cost?

Analyze only the text and operations you need, batch short texts to avoid paying the per-request minimum repeatedly, use asynchronous (batch) inference for custom models where real-time is not required so you avoid always-on endpoint hours, and cache results to avoid reprocessing.

Why do many small Comprehend requests cost more?

Because each request has a minimum charge of 3 units (300 characters), so analyzing many short texts individually pays the minimum each time even for texts shorter than 300 characters. Batching short texts into fewer, larger requests amortizes the minimum and reduces cost.

Do Comprehend custom models cost extra?

Yes. Custom classification and entity models add a training cost and, for real-time inference, an endpoint billed per hour while provisioned, whether or not it serves requests. Use asynchronous batch inference where real-time is not needed to avoid the endpoint hours, and undeploy idle endpoints.

Does running multiple Comprehend operations cost more?

Yes. Each operation, sentiment, entities, key phrases, language detection, bills per text unit separately, so running several on the same text bills for each. Run only the operations the use case needs rather than all of them, to avoid multiplying the per-unit cost.

Does C3X estimate Comprehend cost?

Comprehend cost is driven by text volume, operations, and any custom-model endpoints, which are usage inputs. C3X prices the surrounding infrastructure, and you model your text analysis volume to estimate the per-unit NLP charges.

What to do next

Estimate the infrastructure around your NLP pipeline before you build it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.