AWS Pricing Calculator
Calculate your HAQM Personalize and architecture cost in a single estimate.
Free Trial
Get started with AWS Personalize for free.
For the first two months of using HAQM Personalize, you are offered the following:
Data processing and storage: Up to 20 GB per month.
Training:
- Up to 5 million interactions per month for User-Personalization-v2 and up to 5 million interactions per month for Personalized-Ranking-v2.
- Up to 100 training hours per month for other Custom Recommendation Solutions.
Recommendations:
- Up to 50,000 real-time recommendation requests per month for User-Personalization-v2 and Personalized-Ranking-v2.
- Up to 180,000 real-time recommendation requests per month for other Custom Recommendation Solutions.
Enhanced Custom Recommendation Solutions
HAQM Personalize v2 recipes
HAQM Personalize v2 recipes (User-Personalization-v2 and Personalized-Ranking-v2) use a Transformer-based architecture, making it easy for you to build a wide array of personalization experiences without requiring machine learning expertise.
There are three components of costs to use v2 recipes:
Data ingestion: You are charged per GB of data uploaded to HAQM Personalize. This includes real-time data streamed to HAQM Personalize and batch data uploaded via HAQM Simple Storage Service (S3).
Training: For each model training job, you are charged based on the number of interactions ingested for training. You can ingest interactions via real-time data streaming or S3 batch uploads. If you ingest more interactions than the service quota, you will be charged based on the maximum number of item interactions that are considered by a model during training (3 billions by default).
Inference: You are charged based on the number of recommendation requests for both real-time and batch recommendations. For real-time recommendations, HAQM Personalize charges a minimum of 1 recommendation request transaction per second (TPS) for all active campaigns by default. This minimum 1 TPS charge applies even if you make no requests. You can also provision a higher minimum transaction rate if needed. When the rate of recommendation requests exceeds the minimum provisioned TPS, HAQM Personalize auto-scales to serve your requests and returns to the minimum provisioned TPS when your traffic reduces. You are billed for the greater of minimum provisioned TPS (1 TPS by default) and the actual TPS incurred. Pricing examples 1 and 2 illustrate how the real-time inference charge is calculated.
Pricing table
The pricing table applies when using the following recipes:
User-Personalization-v2
Personalized-Ranking-v2
Pricing
|
Data Ingestion
|
$0.05 per GB of data uploaded to HAQM Personalize
|
|---|---|
|
Training
|
$0.002 per 1,000 interactions ingested for training |
|
Inference
|
$0.15 per 1,000 recommendation requests for both real-time recommendations and batch recommendations |
Pricing examples
Example 1: Custom real-time recommendations
A company uses custom real-time recommendations to generate recommendations for a carousel on their homepage. They upload 200 GB of data in the month, and train their solution twice per week with each training considering 10M interactions ingested. For 10 hours per day, the carousel receives traffic of 36,000 visits per hour. During non-peak hours, the carousel receives less than 3,600 visits per hour, or less than the minimum transaction rate of 1 TPS. As a result, Personalize auto-scales down to the minimum 1 TPS and the customer is billed for 3,600 recommendation requests per hour during this period (1 transaction per second * 3,600 seconds per hour).
The bill for the month for using HAQM Personalize will be:
Data processing and storage charge = 200 GB * $0.05 per GB = $10.00
Solution training charge = 10 million interactions ingested for training $2.00 per 1M interactions 8 trainings per month = $160.00
Inference consumption and charge (real-time inference):
Peak traffic usage: 36,000 recommendation requests 10 hours per day 30 days per month = 10,800,000 recommendation requests
Off-Peak traffic usage: 3,600 recommendation requests 14 hours per day 30 days per month = 1,512,000 recommendation requests
12,312,000 recommendation requests * $0.15 per 1,000 real-time recommendation requests = $1,846.80
Total cost = $10.00 + $160.00 + $1,846.80 = $2,016.80
Example 2: Custom real-time recommendations with variable inference traffic
For simplicity, let us assume the company in Example 1 creates another carousel of recommendations that uses the same amount of data ingestion and training. However, this carousel’s traffic varies more throughout the day. In this example, the customer has provisioned a higher minimum TPS.
Inference consumption and charge: In following table, we walk through a variable traffic scenario and calculate the recommendation requests consumed in a day of usage:
Inference charge calculation
|
Time
|
Time (hours elapsed)
|
Minimum Provisioned TPS
|
Minimum recommendation request transactions per hour (min. Provisioned TPS * 3,600 seconds per hour)
|
Actual Recommendation Requests per Hour
|
Billed Consumption per Hour [max. (minimum, actual)]
|
Total Billed Consumption (Hourly Consumption * Hours)
|
|---|---|---|---|---|---|---|
|
12:00 am - 6:00 pm
|
18
|
30
|
108,000
|
72,000
|
108,000
|
1,944,000
|
|
6:00 pm - 10:00 pm
|
4
|
30
|
108,000
|
144,000
|
144,000
|
576,000
|
|
10:00 pm - 11:00 pm
|
1
|
30
|
108,000
|
18,000
|
108,000
|
108,000
|
|
11:00 pm - 12:00 am
|
1
|
20
|
72,000
|
0
|
72,000
|
72,000
|
|
Total Recommendations Requests per day
|
|
|
|
|
|
2,700,000
|
|
Total Recommendations Requests per month
|
|
|
|
|
|
81,000,000
|
Inference charge: 81,000,000 recommendation requests * $0.15 per 1,000 real-time recommendation requests = $12,150.00
Example 3: Custom batch recommendations
A company uses Custom Recommendations to generate personalized item recommendations for each user in their email marketing campaigns. They ingest 10 GB of data and 5M interactions for training. The company uses a batch inference to generate recommendations for 1 million users. Each recommendation request returns 10 items per user, however the company is only charged for the 1 million requests.
In this case the charges for using Personalize will be:
Data processing and storage charge = 10 GB * $0.05 per GB = $0.50
Solution training charge = 5M interactions ingested for training * $2.00 per 1M interactions = $10.00
Inference charge = 1 million requests * $0.15 per 1,000 real-time recommendation requests = $150.00
Total cost = $0.50 + $10.00 + $150.00 = $160.50
Custom Recommendation Solutions
Real-time recommendations
|
Real-time recommendations
|
Price per 1,000 recommendation requests
|
|---|---|
|
First 72 million requests per month
|
$0.0556 |
|
Next 648 million requests per month
|
$0.0278 |
|
Over 720 million requests per month
|
$0.0139 |
* HAQM Personalize allows you to configure your campaign to return Item metadata with the recommendation request response. You are charged an additional $0.0167 per 1,000 recommendation requests for all campaigns with Item metadata enabled. Note that this additional charge also applies to the minimum provisioned TPS when Item metadata is enabled.
Batch recommendations
For batch recommendations, you are charged for the number of recommendations requested, regardless of the number of results returned. Content Generator uses large language models to generate themes for batch recommendations. You are charged an additional $1 per theme output.
|
Batch recommendations
|
Price per 1,000 recommendations
|
|---|---|
|
First 20 million recommendations per month per eligible Region
|
$0.067
|
|
Next 180 million recommendations per month per eligible Region
|
$0.058
|
|
Over 200 million recommendations per month per eligible Region
|
$0.050
|
Pricing examples
Example 1: Custom real-time recommendations
A company uses custom real-time recommendations to generate recommendations for a carousel on their homepage. They upload 200 GB of data in the month, and train their solution twice per week with each training consuming 15 training hours. For 10 hours per day, the carousel receives traffic of 36,000 visits per hour. During non-peak hours, the carousel receives less than 3,600 visits per hour, or less than the minimum transaction rate of 1 TPS. As a result, Personalize auto-scales down to the minimum 1 TPS and the customer is billed for 3,600 recommendation requests per hour during this period (1 transaction per second * 3,600 seconds per hour).
The bill for the month for using HAQM Personalize will be:
• Data processing and storage charge = 200 GB $0.05 per GB = $10.00
• Solution training charge = 15 training hours 8 trainings per month $0.24 per training hour = $28.80
• Inference consumption and charge (real-time inference)
o Peak traffic usage: 36,000 recommendation requests 10 hours per day 30 days per month = 10,800,000 recommendation requests
o Off-Peak traffic usage: 3,600 recommendation requests 14 hours per day* 30 days per month = 1,512,000 recommendation requests
o 12,312,000 recommendation requests * $0.0556 per 1,000 real-time recommendation requests = $684.55
Total cost = $10 + $28.80 + $684.55 = $723.35
Example 2: Custom real-time recommendations with variable inference traffic
For simplicity, let us assume the company in Example 1 creates another carousel of recommendations that uses the same amount of data ingestion and training. However, this carousel’s traffic varies more throughout the day. In this example, the customer has provisioned a higher minimum TPS.
Inference consumption and charge: In following table, we walk through a variable traffic scenario and calculate the recommendation requests consumed in a day of usage:
Inference charge calculation
|
Time
|
Time (hours elapsed)
|
minProvisioned TPS
|
Minimum recommendation request transactions per hour (min. Provisioned TPS * 3,600 seconds per hour)
|
Actual Recommendation Requests per Hour
|
Billed Consumption per Hour [max. (minimum, actual)]
|
Total Billed Consumption(Hourly Consumption * Hours)
|
|---|---|---|---|---|---|---|
|
12:00 am - 6:00 pm
|
18
|
30
|
108,000
|
72,000
|
108,000
|
1,944,000
|
|
6:00 pm - 10:00 pm
|
4
|
30
|
108,000
|
144,000
|
144,000
|
576,000
|
|
10:00 pm - 11:00 pm
|
1
|
30
|
108,000
|
18,000
|
108,000
|
108,000
|
|
11:00 pm - 12:00 am
|
1
|
20
|
72,000
|
0
|
72,000
|
72,000
|
|
Total Recommendations Requests per day
|
|
|
|
|
|
2,700,000
|
|
Total Recommendations Requests per month
|
|
|
|
|
|
81,000,000
|
Tiers
|
Total recommendation (inference) charge
|
Usage Recommendation Requests (in Tier)
|
Price per 1,000 Real-Time Recommendation Requests
|
Cost ($)
|
|---|---|---|---|
|
Tier 1
|
72,000,000
|
$0.0556
|
$4,003
|
|
Tier 2
|
9,000,000
|
$0.0278
|
$250
|
|
|
|
|
$4,253
|
Pricing examples Continued
Example 3: Custom batch recommendations
A company uses Custom Recommendations to generate personalized item recommendations for each user in their email marketing campaigns. They ingest 10 GB of data and training consumes 50 training-hours. The company uses a batch inference to generate recommendations for 1 million users. Each recommendation request returns 10 items per user, however the company is only charged for the 1 million requests.
In this case the charges for using Personalize will be:
Data processing and storage charge = 10 GB * $0.05/GB =$0.50
Solution training charge = 50 training-hours * $0.24/training-hour = $12
Inference charge = 1 million users * $0.067/1,000 recommendations = $67
Total cost = $0.50 + $12 + $67 = $79.50
Example 4: Custom themed batch recommendations with Content Generator
A company uses Custom Recommendations to generate personalized item recommendations with themes. They ingest 10 GB of data and training consumes 50 training-hours. The company uses a batch inference to generate themed recommendations for 100 seed items. Each recommendation request returns 25 items per seed item. The company will get 100 themes in total.
In this case the charges for using Personalize will be:
Data processing and storage charge = 10 GB $0.05/GB =$0.50
Solution training charge = 50 training-hours $0.24/training-hour = $12
Inference charge = 100 seed items $0.067/1,000 recommendations + 100 themes $1/theme = $100.0067
Total cost = $0.50 + $12 + $100.0067 = $112.5067
Use Case Optimized Recommenders
HAQM Personalize
HAQM Personalize offers use case optimized recommenders that simplify the creation and maintenance of common recommendation solutions. Select the recommenders you wish to use and HAQM Personalize automatically configures the underlying machine learning (ML) models and fully manages their lifecycle. You can select from nine recommenders that offer personalized recommendations for different touchpoints in your user experience.
The following pricing applies when using the following recipes:
aws-ecomm-popular-items-by-view
aws-ecomm-popular-items-by-purchases
aws-ecomm-frequently-bought-together
aws-ecomm-customers-who-viewed-x-also-viewed
aws-ecomm-recommended-for-you
aws-vod-most-popular
aws-vod-because-you-watched-x
aws-vod-more-like-x
aws-vod-top-picks
Data ingestion
You are charged per GB of data uploaded to HAQM Personalize. This includes real-time data streamed to HAQM Personalize and batch data uploaded via HAQM Simple Storage Service (S3).
Data ingestion costs: $0.05 per GB
Recommender Hours
You are charged an hourly rate for each active recommender based on the number of users* in your datasets processed by HAQM Personalize. Each recommender serves a fixed recommendations per hour at no extra cost based on the number of users in your dataset.
|
Users per recommender
|
Price per 100,000 users
|
Free recommendations per hour
|
|---|---|---|
|
First 100,000 users
|
$0.375
|
4,000
|
|
Next 900,000 users
|
$0.045
|
6,000
|
|
Next 9 million users
|
$0.018
|
9,000
|
|
Over 10 million users
|
$0.005
|
14,000
|
* HAQM Personalize allows you to configure your Recommender to return Item metadata in the API response. You are charged an additional $0.1 per hour for Recommenders that are configured to return Item metadata.
Additional recommendations
When the recommendations in an hour exceed the free recommendations for the user tier (see table above), you are charged for the additional recommendations used per hour.
|
Additional recommendations
|
Price per 1,000 recommendations
|
|---|---|
|
First 100,000 recommendations per hour per eligible Region
|
$0.0833
|
|
Next 900,000 recommendations per hour per eligible Region
|
$0.0417
|
|
Over 1 million recommendations per hour per eligible Region
|
$0.0208
|
* HAQM Personalize allows you to configure your Recommender to return Item metadata in the API response. You are charged an additional $0.0167 per 1,000 additional recommendations for all Recommenders with Item metadata enabled.
*The number of users (identified with a ‘user_id’) is calculated as the number of unique users in the union of your ‘Users’ and ‘Interactions’ datasets.
You have the option to specify the minimum throughput for Use Case Optimized Recommenders in recommendations per second (RPS). If the minimum provisioned RPS exceeds the actual recommendations requested per second, the minimum provisioned RPS will count against the free recommendations per hour included in your user tier. If the minimum provisioned RPS causes you to exceed the free recommendations per hour included in your user tier, you will also be charged for additional recommendations. For example, if you set the minimum RPS to 10, you will be charged against the 36,000 recommendations for that hour (3,600 second per hour x 10 RPS) barring the free recommendations per hour in your user tier.
Pricing examples
Example 1: Use case optimized recommenders for a media company
A media company powers three different carousels of recommendations on their app using three use case optimized recommenders. They ingest 200 GB of data in the month and have 2,000,000 users. The carousels each typically see fewer than 9,000 visits per hour; however, there are 140 peak hours per month where they see 39,000 visits per hour.
The bill for the month for using HAQM Personalize will be:
Data processing and storage charge = 200 GB * $0.05 per GB = $10
Recommender Hour charge:
First 100,000 users = $0.375 per hour 720 hours per month 3 recommenders = $810.00
Next 900,000 users = 900,000 users $0.045 per hour/100,000 users 720 hours per month * 3 recommenders = $874.80
Next 1,000,000 users = 1,000,000 users $0.018 per hour/100,000 users 720 hours per month * 3 recommenders = $388.80
Total recommender hour charges = $810.00 + $874.80 + $388.80 = $2,073.60
Additional recommendations charge:
39,000 recommendations per peak hour – 9,000 free recommendations per hour = 30,000 additional recommendations per hour.
30,000 additional recommendations per peak hour $0.0833/1,000 recommendations 140 peak hours * 3 recommenders = $1,049.58
Total cost = $10 + $2,073.60 + $1,049.58 = $3,133.18
Example 2: Use case optimized recommenders for an online retailer
An online retailer uses four use case optimized recommenders to serve product recommendations on their product detail page. They upload 10 GB of data in the month and have 800,000 users. The traffic to these recommenders never exceeds 6,000 visits per hour.
The bill for the month for using HAQM Personalize will be:
Data processing and storage charge = 10 GB * $0.05/GB =$0.50
User charges:
First 100,000 users = $0.375 per hour 720 hours per month 4 recommenders = $1,080.00
Next 700,000 users = 700,000 users $0.045 per hour/100,000 users 720 hours per month * 4 recommenders = $907.20
Total recommender hour charges = $1080.00 + $907.20 = $1,987.20
Additional recommendation charge:
Because the company never exceeds the 6,000 recommendations per hour included with their recommenders there are no additional recommendation charges.
Total cost = $0.50 + $1,987.20 = $1,987.70
User Segmentation
Batch segments (inference)
You are charged for the number of segments requested based on the number of users* in the dataset processed by HAQM Personalize.
|
Users in dataset
|
Price per 1,000 users per segment
|
|---|---|
|
First 100,000 users
|
$0.016
|
|
Next 900,000 users
|
$0.008
|
|
Next 9 million users
|
$0.004
|
|
Next 40 million users
|
$0.001
|
*The number of users (identified with a ‘user_id’) is calculated as the unique users in the union of your ‘Users’ and ‘Interactions’ datasets.
Pricing examples
Example 1: Batch segmentation at an online retailer
A retailer uses batch segmentation to generate lists of users for SMS and in-app messaging campaigns about particular products that are on sale. They run campaigns on 10 products and consider 2,000,000 users for each campaign. They ingest 10GB of data and training requires 50 training-hours.
The bill for using HAQM Personalize for these campaigns will be:
Data processing and storage charge = 10 GB * $0.05/GB = $0.50
Solution training charge = 50 training-hours * $0.24/training-hour = $12.00
Batch segment generation charge, first 100,000 users = 100,000 users $0.016/1,000 users 10 queries = $16.00
Batch segment generation charge, next 900,000 users = 900,000 users $0.008/1,000 users 10 queries = $72.00
Batch segment generation charge, next 1,000,000 users = 1,000,000 users $0.004/1,000 users 10 queries = $40.00
Total cost = $0.50 + $12 + $16 + $72 + $40 = $140.50
Example 2: Batch segmentation at a media company
A media company uses batch segmentation to identify users that would be interested in streaming movies based on attributes of the movies, such as genre, lead actor/actress, and awards won. The company uses the segments of users generated to target their email marketing campaigns. The company has 20 million users that are considered for each campaign. The company uses 650 GB of data, and training requires 1,800 training-hours. They run segmentation on 25 different movie attributes for their campaigns.
The bill for the month for using HAQM Personalize will be:
Data processing and storage = 650 GB * $0.05/GB =$32.50
Solution training charge = 1,800 training-hours * $0.24/training-hour = $432.0
Inference charge, first 100,000 users = 100,000 users $0.016/1,000 users 25 queries = $40
Batch segment generation charge, next 900,000 users = 900,000 users $0.008/1,000 users 25 queries = $180
Batch segment generation charge, next 9 million users = 9,000,000 users $0.004/1,000 users 25 queries = $900
Batch segment generation charge, next 10 million users = 10,000,000 users $0.001/1,000 users 25 queries = $250
Total cost = $32.50 + $432 + $40 + $180 + $900 + $250 = $1,834.50