Amazon SageMaker pricing: every component, with real numbers.
SageMaker AI has no licence fee. You pay per second for the compute you run, plus storage and the managed features you use. Two choices decide most of the bill: the instance (a top GPU costs hundreds of times a small CPU) and the endpoint mode (an always-on endpoint bills while idle; serverless and batch do not). This page lists each meter, AWS list prices for common instances, Savings Plans, two worked examples and the levers that cut cost.
- SageMaker AI (formerly Amazon SageMaker) has no licence fee. You pay per second for training jobs, endpoints, notebooks and processing, plus Feature Store, labeling and storage. Compute dominates almost every bill, and the GPU you pick and the endpoint mode you choose move it more than anything else.
- The most common surprise bill is an always-on real-time endpoint or an idle notebook left running: they bill around the clock. Serverless inference and batch transform scale to zero. At AWS list prices (us-east-1), one ml.g5.xlarge real-time endpoint costs about $1,028 a month left on; the same model scored nightly by batch transform costs about $42.
- SageMaker Savings Plans cut compute cost by up to 64% for a 1- or 3-year commitment, and managed Spot cuts training cost. AWS credits apply to SageMaker usage. A vetted AWS partner can file POC or incremental workload (IW) funding for ML work on AWS: for startups often $50K to $150K. Depends on the workload; AWS approves.
IHow SageMaker billing works
Pure usage pricing: no licence fee, compute billed per second, and a separate meter for each component.
There is no subscription or seat fee. Every dollar traces to a resource you turned on. AWS describes on-demand pricing as having no minimum fees and no upfront commitments, and compute is billed per second, which is why short training jobs are cheap: you pay for the seconds the cluster existed.
The meters: training compute (per instance-second while a job runs); inference compute (instances behind real-time or async endpoints, compute per request for serverless, or job time for batch transform); notebook and Studio compute; Feature Store (writes, reads, storage); Data Wrangler and processing; Ground Truth labeling; and storage and data transfer.
Compute dominates. Training hours and always-on endpoint hours are the big lines; the rest is small until you reach large scale.
Prices on this page are AWS list prices for us-east-1 on-demand from the AWS Price List, checked October 10, 2026. Prices vary by Region and change; confirm on the AWS SageMaker AI pricing page before you budget.
IIPricing by component
What triggers each meter. The pattern: things that run all the time cost the most; things that run briefly are cheap.
- Training jobs: Per instance-second for every instance in the job, from start to finish. Cost = hourly rate ÷ 3,600 × seconds × instance count. Managed Spot training lowers this for jobs that can be interrupted.
- Real-time endpoints: Per instance-hour for as long as the endpoint exists, whether or not requests arrive. The number one source of surprise bills. Auto-scaling adds instances under load.
- Serverless inference: Billed for compute used per request (memory size × duration) plus requests. Scales to zero when idle. Cheaper than real-time for spiky traffic, with occasional cold starts.
- Asynchronous inference: Per instance-time while processing a queue; can scale to zero between bursts. For large payloads and long inferences.
- Batch transform: Per instance-second of a short job that scores a dataset. Nothing bills between runs. Usually the cheapest way to score offline.
- Studio and notebook compute: Per instance-hour while the notebook or app runs. An idle but running notebook keeps billing, so set auto-shutdown. The domain itself is free.
- Feature Store, Data Wrangler, Ground Truth: Feature Store bills writes, reads and storage; Data Wrangler bills processing time; Ground Truth bills per labeled object plus any human workforce.
- Storage and data transfer: S3 for datasets and model artifacts at S3 rates, attached volumes, and standard data-transfer charges.
Real-time endpoints and notebooks bill while idle; serverless, async and batch transform scale to zero. Auto-shutdown notebooks, delete test endpoints and default to serverless or batch until steady traffic justifies real-time.
IIIInstance prices: AWS list prices for common types
Because compute dominates, the instance is the lever. At us-east-1 list prices, an ml.p5.48xlarge costs over 500 times an ml.m5.large per hour.
| Instance | Accelerator | Price per hour | Typical use |
|---|---|---|---|
| ml.t3.medium (notebook) | None | $0.05 | Notebooks, light work |
| ml.m5.large | None | $0.115 | Light inference, small tabular training |
| ml.c5.xlarge | None | $0.204 | CPU training, batch scoring |
| ml.m5.xlarge | None | $0.23 | General CPU training and hosting |
| ml.g4dn.xlarge | 1 NVIDIA T4 | $0.736 | Entry GPU inference |
| ml.inf2.xlarge (hosting) | AWS Inferentia2 | $0.99 | Lower-cost high-volume inference |
| ml.g6.xlarge | 1 NVIDIA L4 | $1.127 | Entry GPU inference and fine-tuning |
| ml.g5.xlarge | 1 NVIDIA A10G | $1.408 | Small-model training, inference, fine-tuning |
| ml.trn1.2xlarge (training) | AWS Trainium | $1.54 | Lower-cost training |
| ml.g5.12xlarge | 4 NVIDIA A10G | $7.09 | Mid-size training, multi-GPU inference |
| ml.p4d.24xlarge | 8 NVIDIA A100 | $25.25 | Large-model training |
| ml.p5.48xlarge | 8 NVIDIA H100 | $63.30 | Foundation-model and large distributed training |
IVSavings Plans, Spot and the Free Tier
Three ways to pay less than on-demand.
SageMaker Savings Plans. Commit to a set amount of compute spend per hour for 1 or 3 years. AWS says they cut costs by up to 64%. They apply automatically across eligible usage (Studio and notebook instances, processing, Data Wrangler, training, real-time inference and batch transform) regardless of instance family, size or Region. You pay the commitment even if usage drops, so size it to your steady baseline, not your peak.
Managed Spot training. Training jobs run on spare AWS capacity at a lower price, with checkpointing so an interrupted job resumes. Good for development and non-urgent training. Not for production endpoints.
Free Tier. For the first 2 months after your first SageMaker AI resource, AWS lists 250 hours of ml.t3.medium notebooks, 50 hours of m4.xlarge or m5.xlarge training, 125 hours of m4.xlarge or m5.xlarge real-time inference and 150,000 seconds of serverless inference each month, among others.
Savings Plans lower the rate; AWS credits pay eligible usage on the bill. AWS promotional credit terms say credits do not pay Savings Plans or Reserved Instance upfront fees, so read the terms before you commit while you are spending credits.
VTwo worked examples
Training once, and hosting around the clock, using the us-east-1 list prices above.
Example A: train a model once
Scenario: fine-tune a mid-size model on one ml.p4d.24xlarge (8 A100 GPUs) for a 10-hour run.
Compute: $25.25 per hour × 10 hours = about $253 on-demand. Managed Spot lowers this; the saving depends on spare capacity at the time.
Storage: the dataset and model artifact in S3 cost cents to a few dollars.
The point: training is a spike. You pay for those hours, then nothing. Ten development runs cost about $2,525 on-demand, which is the experimentation budget AWS credits can cover.
Example B: host an endpoint around the clock
Scenario: one ml.g5.xlarge real-time endpoint ($1.408 per hour) running for a month (730 hours), no auto-scaling.
Compute: $1.408 × 730 ≈ $1,028 a month, whether it serves a million requests or none.
The same model, different mode: if the work is offline, batch transform for one hour a night is 30 × $1.408 ≈ $42 a month. If traffic is spiky, serverless bills only the compute used.
Training cost depends on instance size and run length. Hosting cost depends on endpoint mode far more than on the model. Get those two right and you have handled most of the bill.
VICost levers, in order of impact
Work down this list. The top items usually dwarf the bottom ones.
- 1. Kill idle endpoints and notebooks: Auto-shutdown notebooks, delete test endpoints when an experiment ends and audit monthly for forgotten ones.
- 2. Match endpoint mode to traffic: Serverless for spiky online traffic, batch transform for offline scoring, async for large or slow requests, real-time only for steady low-latency traffic.
- 3. Use Spot for training: Managed Spot with checkpoints for development and non-urgent runs.
- 4. Right-size the instance: Do not train on 8 H100s what fits on one A10G, and do not host on a GPU what runs fine on CPU.
- 5. Try AWS silicon: Trainium for training and Inferentia for inference, through the Neuron SDK. Migration takes work; savings grow with scale.
- 6. Buy a Savings Plan for the baseline: Once usage is steady, commit the baseline for up to 64% off. Size to baseline, not peak.
- 7. Auto-scale real-time endpoints: Run the minimum off-peak and add capacity only under load.
- 8. Fund it with AWS credits: Credits pay eligible SageMaker usage until they run out. See the next section for who files what.
VIIHow AWS credits and partner funding pay for SageMaker
AWS credits apply to SageMaker compute, storage and features like any other AWS service. Two kinds of AWS money fit, and different people apply for each.
Partner-filed funding (the bigger route for most startups). A vetted AWS consulting partner can file proof of concept (POC) funding to test an ML workload, migration (MAP) funding if you are moving ML work to AWS from another cloud, or incremental workload (IW) funding if you already run on AWS and are launching new work. For startups this is often $50K to $150K. Depends on the workload; AWS approves. The partner files it, not you.
AWS Activate credits (you apply yourself). Activate Founders is $1,000 to start, up to $5,000; you apply on the AWS site. Activate Portfolio is up to $200,000 with the Org ID from a VC, accelerator or angel group that is an Activate Provider. A consulting partner cannot file Activate for you.
CloudRoute matches you with one vetted AWS partner who can cut the SageMaker bill (Spot, right-sizing, endpoint mode) and file the partner funding when the work qualifies. Startups pay CloudRoute nothing. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.
| Funding | Who files it | Amount | When it fits |
|---|---|---|---|
| Migration Acceleration Program (MAP) | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You are not on AWS yet and want to migrate. |
| Proof of concept (POC) funding | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You want to test something, whether or not you are on AWS today. |
| Incremental workload (IW) funding | Your matched AWS partner | $50K to $150K for startups. Depends on the workload; AWS approves. | You are already on AWS and are launching new work: a new AI feature, a new customer or a new contract. |
| AWS Activate Founders | Founder applies | $1,000 to start, up to $5,000 | Self-funded startups with no investor or accelerator behind them. |
| AWS Activate Portfolio | Founder applies | up to $200,000 | Startups backed by a VC firm, accelerator or angel investor that is an AWS Activate Provider. |
VIIIThe pricing mistakes that cost the most
Most SageMaker overspend comes from a short list of repeat mistakes.
- Leaving a real-time endpoint running: An ml.g5.xlarge left up for a month is about $1,028 at us-east-1 list price. Delete endpoints when you are done.
- Defaulting to real-time: Using the always-on mode when traffic is spiky or offline. Match the mode to the traffic.
- Training on a bigger GPU than needed: An ml.p5.48xlarge costs about 45 times an ml.g5.xlarge per hour. Right-size first.
- Skipping Spot for interruptible training: Paying on-demand for development runs that could resume from a checkpoint.
- Sizing a Savings Plan to peak: You pay the commitment whatever you use. Size it to the baseline.
- Forgetting Studio notebooks: Idle notebook compute bills per hour. Turn on auto-shutdown.
Same model, four ways to serve it
The endpoint mode is the biggest lever in SageMaker serving. Monthly figures use the ml.g5.xlarge us-east-1 list price ($1.408 per hour, checked October 10, 2026).
| Endpoint mode | Billing basis | Bills when idle? | Cold starts? | Illustrative monthly cost | Best when |
|---|---|---|---|---|---|
| Real-time | Per instance-hour, always on | Yes | No | About $1,028 (one instance, 730 hours) | Steady, latency-sensitive traffic |
| Serverless | Compute used per request | No, scales to zero | Occasional | A fraction of real-time if busy part of the day | Spiky or intermittent traffic |
| Asynchronous | Instance time while busy | No, scales to zero | Minimal | Proportional to busy time | Large payloads, long inferences |
| Batch transform | Job instance time | No, short job | N/A | About $42 (one hour a night) | Offline, scheduled scoring |
How a match works (example)
- You tell us about the work
You fill in a short form: what you run today, what you want to move, test or build, your stage and your region.
- We pick one partner
We read your answers and pick one vetted AWS consulting partner that fits the work, the stack and the region. We reply within 24 hours; your matched partner reaches out within 5 business days.
- The partner checks POC funding
Your matched partner scopes the test with you. Then it checks whether the test qualifies for POC (proof of concept) funding. A POC can fit whether or not you run on AWS today.
- The partner files, AWS decides
If the work qualifies, the partner files the funding request with AWS. AWS reviews it and decides. The amount depends on the workload.
- You pay nothing
CloudRoute is free for startups. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.
AWS Activate credits are separate, and you apply for them yourself. Use Activate Founders if no investor backs you. Use Activate Portfolio with the Org ID or invite from your VC, accelerator, incubator or angel group if it is an Activate Provider. A consulting partner cannot apply for Activate on your behalf.
Common questions
How much does Amazon SageMaker cost?
Why is my SageMaker bill so high?
What is the cheapest way to host a model on SageMaker?
How do SageMaker Savings Plans work?
Is there a SageMaker free tier?
Is SageMaker or Bedrock cheaper?
Can AWS credits pay for SageMaker?
Cut the SageMaker bill, and get AWS money behind the work
CloudRoute matches you with one vetted AWS partner who cuts SageMaker spend (Spot, right-sizing, endpoint mode, Savings Plans). When the work qualifies, the partner files POC or incremental workload (IW) funding, and AWS decides. Startups pay CloudRoute nothing.