Updated 22 Oct 2026 • 7 mins read
SageMaker Pricing
Amazon SageMaker pricing is a free console over a dozen instance-hour meters, each priced above the equivalent EC2 instance, with endpoints billing around the clock whether or not they serve traffic. This guide gives every rate in US dollars, the premium over EC2, when serverless beats an endpoint, the optimization levers, and how to see ML spend per team and model.
Amazon SageMaker pricing starts with a number that is true and misleading: Studio is free. Everything you do inside it is not. A notebook, a processing job, a training job, a real-time endpoint, a batch transform, and a HyperPod node each bill by the instance-hour on their own meter, at ml-prefixed rates 15 to 40 percent above the identical EC2 instance, and the meter that runs longest is the one nobody is watching. An inference endpoint deployed for a demo and never deleted bills every hour of every day whether it serves a request or not. On an ml.g5.xlarge that is $1,028 a month. Most ML teams have several.
This guide gives every SageMaker meter and rate in US dollars as of October 2026, the premium each carries over EC2, the serverless-versus-endpoint crossover, the training rates including the October 7 reset for P5 and P6 capacity, the optimization levers in order of impact, and the step most ML estates skip: attributing the result to the teams and models that caused it, because SageMaker spend arrives as a line inside the AWS bill that no model is named on. Rates are from the official SageMaker AI pricing page; the wider practice is in our guide to FinOps for AI.
The short answer: Amazon SageMaker charges nothing for Studio and bills every compute surface by the instance-hour: notebooks, processing, training, real-time and asynchronous endpoints, batch transform, and HyperPod, on ml instance types priced 15 to 40 percent above the equivalent EC2 instance. In US East an ml.m5.xlarge is $0.23 an hour (EC2: $0.192), an ml.g5.xlarge endpoint $1.408 (EC2: $1.006), and an ml.p5.48xlarge about $63.30 on demand (EC2: $55.04), with Flexible Training Plan rates reset on October 7, 2026 to $54.92 for P5 and $130.71 to $148.54 for the Blackwell P6 instances. Serverless inference bills $0.20 per million requests plus memory-seconds and wins below roughly 700,000 requests a month for a typical model; endpoints bill around the clock and win at steady high traffic. Managed Spot Training cuts training 60 to 90 percent; ML Savings Plans cut notebooks, training, and real-time endpoints by up to 64 percent but not serverless or asynchronous inference. The largest waste in most estates is idle endpoints and notebooks, found only by inventory.
Key Takeaways
- Studio is free; a dozen instance-hour meters are not, and each carries a 15 to 40 percent premium over the same EC2 instance.
- Endpoints and notebooks bill whether or not they are used; one forgotten ml.g5.xlarge endpoint is $1,028 a month.
- Serverless inference wins at low and variable traffic, endpoints at steady high traffic; the crossover is a calculation, not a default.
- Managed Spot Training (60 to 90 percent off) and ML Savings Plans (up to 64 percent) are large and cover different meters.
- SageMaker spend lands in the AWS bill with no model or team on it; attribution is the step that makes the levers recur.
Every SageMaker Meter
US East (N. Virginia), October 2026.
| Meter | How it bills | Example rate | Notes |
|---|---|---|---|
| Studio, Unified Studio | Free | $0 | The console; compute underneath is metered |
| Notebooks, JupyterLab, Code Editor spaces | Instance-hour while the space runs | ml.t3.medium about $0.05; ml.m5.xlarge $0.23 | Bills while open, including overnight |
| Processing jobs, Data Wrangler | Instance-hour for the job | ml.m5.xlarge $0.23; ml.m5.4xlarge $0.922 | Per job |
| Training jobs | Instance-hour times instance count | ml.m5.xlarge $0.23; ml.g5.xlarge about $1.41; ml.p5.48xlarge about $63.30 | Managed Spot Training 60 to 90 percent off |
| Flexible Training Plans (reserved GPU capacity) | Instance-hour, reserved in advance | ml.p5.48xlarge $54.92; ml.p6-b200.48xlarge $130.71; ml.p6-b300.48xlarge $148.54 | Rates reset October 7, 2026; dynamic with demand |
| Real-time inference endpoints | Instance-hour per endpoint instance, deployed or not | ml.m5.large $0.115; ml.m5.xlarge $0.23; ml.g5.xlarge $1.408 | The idle-endpoint trap; autoscaling helps, scale-to-zero does not apply |
| Serverless inference | $0.20 per million requests plus about $0.0000166 per GB-second of memory | 4 GB model at 500 ms: about $0.00024 per request | 150,000 seconds free a month; scales to zero |
| Asynchronous inference | Instance-hour, can scale to zero | As real-time rates | For large payloads and long inference |
| Batch transform | Instance-hour for the job | As training rates | For offline scoring |
| HyperPod clusters | Instance-hour per node while the cluster runs | ml.g5.24xlarge $10.18 | For large-scale distributed training |
| ML storage (EBS for notebooks and endpoints) | Per GB-month | About $0.14 | Persists with the resource |
| Savings Plans | 1- or 3-year hourly commitment | Up to 64 percent off | Covers notebooks, processing, training, real-time endpoints; not serverless or async |
Serverless or Endpoint: The Crossover
A real-time endpoint bills its instance every hour; serverless bills each request. The right choice is arithmetic. Take a model configured at 4 GB that takes 500 milliseconds per inference: serverless costs $0.20 per million requests plus 4 GB times 0.5 seconds times $0.0000166, about $0.00024 per request. An always-on ml.m5.xlarge endpoint costs $168 a month regardless. At 100,000 requests a month serverless is about $24 and wins by $144; at 500,000 it is about $120 and still wins; at 700,000 the two are even; at 2 million serverless is about $480 and the endpoint wins by $312. Shape changes the crossover, since heavier models and longer inference move it lower, but the method does not: count the requests before choosing the meter, and revisit as traffic grows.
A Worked Example
A mid-size ML team runs SageMaker for ten models. Its monthly bill, read from the AWS invoice, is $14,200, in one line.
Broken down, it is: five Studio spaces on ml.g5.xlarge left running around the clock, $5,140; a weekly fine-tuning job on ml.p4d-class capacity at on-demand rates, about $2,260 that Managed Spot would cut to about $700; six real-time endpoints on ml.g5.xlarge at $1,028 each, $6,170, of which three belong to models retired two quarters ago and one serves under 50,000 requests a month; and $630 of storage, processing, and transfer. Stopping the spaces outside working hours saves about $3,900. Moving the fine-tuning job to Spot saves about $1,560. Deleting the three retired endpoints saves $3,080, and moving the low-traffic one to serverless saves about $1,000. The bill falls from $14,200 to about $4,700, a 67 percent reduction, with the same models in production, and three-quarters of the saving came from resources nobody was using. The invoice could not show that; an inventory with an owner on each endpoint could.
How to Cut SageMaker Costs, in Order
1. Find and delete idle endpoints and notebooks
List every endpoint with its invocation count and every Studio space with its last activity. Delete endpoints for retired models, and set notebook spaces to stop automatically after idle periods. This is the first and usually the largest saving.
2. Choose the inference meter by request volume
Serverless for low and bursty traffic, asynchronous for large payloads, real-time endpoints with autoscaling for steady load. Multi-model endpoints can serve several low-traffic models from one instance.
3. Use Managed Spot Training for anything checkpointable
Sixty to ninety percent off training compute, with SageMaker handling interruption and resume from checkpoints. Most fine-tuning and experimentation qualifies.
4. Right-size training and endpoint instances
GPU utilization during training and endpoint CPU and memory tell you whether the instance is doing the work. A g5 endpoint serving a small model at 8 percent GPU is a CPU endpoint waiting to happen.
5. Commit the steady remainder
ML Savings Plans at up to 64 percent for notebooks, training, and real-time endpoints that run every hour, sized after steps 1 to 4, and Flexible Training Plans for large reserved GPU runs.
6. Count the services around the model
S3 storage and requests for datasets and artifacts, data transfer, and the feature store all bill separately and are usually left out of the ML budget.
7. Attribute spend to teams and models
SageMaker spend lands in the AWS bill as a service line with no model on it. Until each endpoint, job, and space carries a team and a model, the six levers above are applied once and the waste returns, which is the pattern our guide to AI unit economics describes for tokens and that applies equally to GPU hours.
How Opslyft Does This
The SageMaker problem is that ML spend hides inside the AWS bill: one service line, a dozen meters behind it, and no endpoint, job, or team named on any of them. Explorer reads SageMaker cost from the bill and shows it by meter, endpoint, training job, notebook space, and team, beside the S3, EC2, and data services the same ML work uses and beside the token spend from Bedrock and direct AI providers, tracked over time, so the retired model's endpoint is a line with a name on it rather than a share of $14,200. Canvas puts ML spend per team on the engineering view and per model and product on finance's, from the same numbers.
Dimension Studio attributes shared training clusters, multi-model endpoints, and the S3 datasets several teams read to the models and products behind them by rule, which is what makes cost per model, and cost per prediction for a model that serves a product, reportable. Atlas inventories every endpoint, space, and job with an owner and a cost, flags endpoints with no invocations and spaces with no recent activity, and its recommendations weigh purpose and criticality, so a production endpoint and a demo endpoint at the same instance size get different advice. Monitor catches a training job that ran three times longer than usual, an endpoint that scaled up and stayed there, or a space left running over a weekend, with the cause, routed to the owning team's Slack channel; Tickets turns the deletion, the Spot migration, or the serverless move into an assigned item in Jira, ServiceNow, or Asana tracked to a measured saving; and Iris answers which models cost the most per team this month in plain English.
That attribution, across six platforms and dozens of shared services, is what gave Disprz 100 percent automated allocation and 40 hours a month back, and it is the same discipline our FinOps for AI guide sets out for every kind of AI spend.
The Bottom Line
SageMaker pricing is a free door into a building where every room bills by the hour at a premium, and the rooms nobody is in are the expensive ones. Inventory the endpoints and spaces first and delete or stop what is idle; choose serverless or endpoints by counting requests; put training on Spot; right-size; commit the steady remainder; count the S3 and transfer around the model; and attribute every meter to a team and a model so the savings hold. Done in that order, ML estates commonly cut SageMaker spend by half or more without touching a production model. Seeing that spend per team and per model, next to the rest of the AI and cloud bill, is what Opslyft's cost visibility is built for; for the GPU side of the same question, see our GPU cost allocation guide.
FAQs
SageMaker Studio itself is free; everything you run inside it bills by the instance-hour on a separate meter: notebooks and JupyterLab spaces, processing jobs, training jobs, real-time and asynchronous inference endpoints, batch transform, and HyperPod cluster nodes. Each uses ml-prefixed instance types priced 15 to 40 percent above the equivalent EC2 instance; an ml.m5.xlarge is $0.23 an hour in US East against $0.192 for an m5.xlarge. Serverless inference bills per request and per memory-second, storage bills per GB-month, and ML Savings Plans discount training and real-time endpoints by up to 64 percent.
Real-time endpoints left running. An endpoint bills every hour it is deployed whether it serves a million requests or none, so a single ml.g5.xlarge endpoint at $1.408 an hour costs about $1,028 a month idle, and three ml.m5.xlarge endpoints from finished experiments cost about $504 a month serving nothing. Studio notebooks left running overnight are the second trap for the same reason.
By the instance-hour at ml rates: about $0.23 an hour on ml.m5.xlarge, $1.41 on ml.g5.xlarge, $63.30 on-demand for ml.p5.48xlarge with eight H100s, and $130 to $149 on the Blackwell-based ml.p6 instances under Flexible Training Plans as of October 7, 2026. Managed Spot Training cuts training cost by 60 to 90 percent for interruptible jobs with checkpointing, and ML Savings Plans by up to 64 percent for steady usage.
At low or intermittent traffic. Serverless bills $0.20 per million requests plus about $0.0000166 per GB-second of configured memory, with 150,000 seconds free a month, so a model at 4 GB taking 500 milliseconds costs roughly $0.00024 per request; a million requests a month is about $240 against $168 for an always-on ml.m5.xlarge endpoint, but at 100,000 requests serverless is about $24 against the same $168. Below roughly 700,000 requests a month at that shape, serverless wins; steady high traffic belongs on an endpoint.
No. Amazon SageMaker Savings Plans, up to 64 percent off for a one- or three-year hourly commitment, apply to notebooks, processing, training, and real-time endpoints across instance families and regions, but not to serverless inference, asynchronous inference, or several newer meters. Size the commitment to steady usage on the covered meters only, after idle endpoints have been removed.
Because the ml instance rates carry a managed-service premium, about 20 percent on CPU instances, 40 percent on G5 GPUs, and 15 percent on P5, in exchange for managed training orchestration, endpoint hosting, autoscaling, and the Studio tooling. For steady, high-utilization inference the premium can exceed the operations it saves; for experimentation, training orchestration, and variable serving it usually does not.
Opslyft reads SageMaker spend from the AWS bill and shows it in Explorer by endpoint, training job, notebook, and team, beside the S3, EC2, and data services the same ML work uses. Dimension Studio attributes shared endpoints and training clusters to the models, teams, and products behind them, Atlas inventories every endpoint and notebook with an owner and flags the idle ones, and Monitor catches a training job or endpoint whose cost jumped with the cause, so ML spend stops being one line inside the AWS bill.