Amazon Comprehend Review 2026

• ✓50,000 units per month for each NLP API (entity recognition, sentiment, key phrases, language detection, syntax)
• ✓5,000 units per month for custom classification and custom entity recognition inference
• ✓Available for the first 12 months after AWS account creation
• ✓Full access to all standard API features during trial period

• ✓Entity recognition, sentiment analysis, key phrase extraction at $0.0001 per unit (1 unit = 100 characters, minimum 3 units per request)
• ✓Language detection at $0.0001 per unit
• ✓Syntax analysis at $0.00005 per unit
• ✓PII detection and redaction at $0.0001 per unit
• ✓Volume discounts available at higher tiers (10M+ units)

• ✓Custom classification model training at $3 per hour
• ✓Custom entity recognition model training at $3 per hour
• ✓Model endpoint hosting at $0.50 per hour ($360/month always-on)
• ✓Asynchronous custom inference also available per-unit
• ✓Model management and versioning included

✅Pros

• •Fully managed with no infrastructure to provision — scales automatically from a single document to millions via asynchronous batch jobs on S3
• •Generous 12-month free tier covering 50,000 units per month across all standard APIs, making it easy to prototype and evaluate without upfront cost
• •Deep AWS ecosystem integration with native S3, Lambda, CloudWatch, KMS, IAM, and 200+ other AWS service connections for building end-to-end pipelines
• •Custom classification and entity recognition models can be trained without ML expertise using simple labeled CSV or augmented manifest files
• •Comprehend Medical provides HIPAA-eligible medical NLP with ontology linking to ICD-10-CM, RxNorm, and SNOMED CT for healthcare use cases
• •Built-in PII detection and redaction supporting 30+ entity types enables compliance workflows without custom regex or third-party tools

❌Cons

• •Language support is uneven — many features only support English and a subset of other languages, limiting usefulness for global multilingual deployments
• •Accuracy can vary significantly by domain; pre-trained models perform best on general-purpose text and may require custom training for specialized terminology
• •Custom model endpoint pricing at $0.50/hour ($360/month) creates ongoing costs even during idle periods, making it expensive for intermittent or low-traffic workloads
• •Vendor lock-in to AWS ecosystem — migrating NLP pipelines to another provider requires rewriting integrations, retraining custom models, and rearchitecting data flows
• •No on-premises or edge deployment option; all processing requires sending data to AWS cloud endpoints, which may conflict with data residency or air-gapped requirements