Amazon EC2 pricing is built around four purchase models: On-Demand, Savings Plans, Reserved Instances, and Spot Instances. You pick a model, then pay for compute based on the instance type, the region you run in, and how long the instance runs. 

EC2 compute is billed in one-second increments with a 60-second minimum for Linux instances, so a server you run for 90 seconds is billed for 90 seconds, not a full hour (AWS billing and purchasing options).

Here is how the four models compare:

The cheapest model depends on workload shape: steady workloads favor Savings Plans or Reserved Instances, bursty workloads favor On-Demand, and interruption-tolerant workloads favor Spot. Most teams combine them.

How AWS EC2 pricing is structured

Your EC2 price for any given server is the product of three things: the instance type (how much CPU, memory, and networking it provides), the region it runs in (the same instance can cost more in São Paulo than in Virginia), and the pricing model you commit to. New to the underlying service? Start with what Elastic Compute Cloud is, then come back here for the pricing layer.

This guide covers how each model is priced and how to estimate it. For a line-by-line breakdown of everything that lands on your monthly EC2 bill, including storage, data transfer, and idle resources, see our dedicated Amazon EC2 costs guide.

The four EC2 pricing models

On-Demand: the no-commitment baseline

On-Demand is the default. You launch an instance, you pay the published per-hour (billed per-second) rate, and you stop paying the moment you terminate it. There is no contract and no upfront payment.

Real rates make this concrete. In us-east-1 (Linux), representative On-Demand rates run roughly:

Rates are illustrative and vary by region and OS. Check live figures on the EC2 On-Demand pricing page before you model. The takeaway that practitioners stress: On-Demand is the most expensive way to run a steady compute, but the cheapest way to run something you will switch off soon. 

A server that genuinely runs 24/7 is the textbook candidate to move off On-Demand and onto a commitment.

Savings Plans: commit to a dollar amount, not a server

Savings Plans flip the commitment model. Instead of reserving a specific instance, you commit to a steady spend (for example, $10 per hour of compute) for one or three years, and AWS discounts everything up to that level. Compute Savings Plans reduce costs by up to 66% and apply automatically across instance families, sizes, regions, and even to Fargate and Lambda. EC2 Instance Savings Plans go up to 72% in exchange for locking to one instance family in one region (AWS Compute Savings Plans).

The two variants trade flexibility against discount depth, and the best-case effective rate on an m5.large shows how far the discount reaches:

The m5.large figures are derived from the On-Demand rate above and the maximum discounts cited here. This is the model the community now reaches for first. As practitioners summarize it on Quora and AWS forums, Savings Plans decouple the commitment from the specific instance type, which removes the single biggest complaint people had about Reserved Instances: getting stuck on a machine type they later outgrew. 

A common rule of thumb shared by FinOps teams is to cover 70 to 80% of your steady baseline with a Compute Savings Plan and leave the rest flexible. If you are weighing the two commitment models directly, our Savings Plans vs Reserved Instances breakdown runs the full comparison.

Reserved Instances: commit to a specific machine

Reserved Instances (RIs) are the original commitment model: you reserve a specific instance type in a specific region for one or three years and receive a discount in return. They come in two main flavors, and the discount depends on which you choose:

Payment options (all upfront, partial upfront, no upfront) trade cash flow against discount depth, with all-upfront giving the lowest effective rate (EC2 Reserved Instance pricing). 

A grounded, real-world comparison from a practitioner write-up: for 10 m5.large-class instances, On-Demand ran about $8,409 per year, while an all-upfront Standard RI cut that roughly 40% to about $5,050 per year, and Savings Plans landed close behind with far more flexibility (Oneuptime EC2 commitment comparison). The lesson teams repeat: RIs win on raw discount for truly static capacity, but most workloads are not static enough to justify giving up flexibility.

Spot Instances: spare capacity at the deepest discount

Spot Instances sell you the unused capacity AWS has on hand at up to 90% off On-Demand, with one catch: AWS can reclaim the instance with a two-minute warning when it needs the capacity back.

The caveat the headline 90% hides: that 90% is a ceiling, not the norm. Real Spot savings depend on the instance type, the region, and how much spare capacity AWS has at the time, so most workloads land below the maximum. AWS publishes the actual trailing-30-day savings and interruption rate for every instance type in its Spot Instance Advisor, which is the authoritative place to check what a specific instance will save.

Spot shines for fault-tolerant work: batch processing, CI/CD runners, big-data jobs, ML training, and stateless web tiers. It is the wrong tool for a stateful database that cannot survive a two-minute eviction. For patterns that hold up in production, see maximizing cloud ROI with Spot Instances.

EC2 pricing by instance type

Beyond the pricing model, the instance family you choose moves the rate as much as anything. The price difference between two equally large instances can be substantial because they target different workloads:

• General purpose (t, m series): balanced CPU-to-memory, the default for web and app servers.

• Compute optimized (c series): higher CPU per dollar, for batch and high-traffic front ends.

• Memory optimized (r, x series): more RAM per dollar, for in-memory caches and large databases.

• Accelerated and GPU (p, g series): the most expensive, for ML and rendering, covered in detail below.

Processor architecture is the other lever. AWS Graviton (Arm-based) instances are priced below comparable x86 instances, with reported savings of up to roughly 20%, and AMD-based instances run up to about 10% cheaper than their Intel equivalents. If your workload runs on Arm, Graviton is often the single easiest price cut available.

When people search for EC2 server pricing, they usually want the rate for one specific instance. The most-requested general-purpose size is the t3.medium, a small burstable server that costs about $0.0416 per hour On-Demand in us-east-1 (Linux), or roughly $30 per month if left running. Burstable t-family instances are billed on a CPU-credit model: they run cheaply at a baseline CPU level and draw credits when they burst above it.

GPU instance pricing (g4dn, A100, H100)

GPU servers are the most expensive EC2 instances by a wide margin, and their rates are what most "AWS GPU pricing" searches are after. A few representative On-Demand rates in us-east-1 (Linux):

Rates from Wring's AWS GPU instance pricing guide. Two practical notes the rates do not show. First, the g4dn family has no g4dn.large size; it starts at g4dn.xlarge, so a search for "g4dn.large pricing" should point to the xlarge as the entry rate. 

Second, GPU instances are where Spot and commitment discounts pay off most in absolute dollars: 70% off a $98 per hour H100 server saves far more than the same percentage off a general-purpose box, which is why GPU workloads are the first place to apply GPU cost optimization.

Is there an AWS EC2 free tier

Yes. New AWS accounts get 750 hours per month of t2.micro or t3.micro Linux usage for the first 12 months, enough to keep one small instance running continuously at no charge (EC2 pricing overview). It is built for learning and light prototyping. Production workloads outgrow it quickly, but it is the cheapest way to get hands-on before you model real spend.

How to estimate your EC2 pricing

Before launching anything, model the cost. The fastest path is the official calculator:

1. Open the AWS Pricing Calculator and add an Amazon EC2 estimate.

2. Choose your region, instance type, and operating system, then set expected hours per month (or quantity of always-on instances).

3. Select the pricing model (On-Demand, Savings Plan, or Reserved) to compare the discounted total against the On-Demand baseline.

The single most useful comparison the calculator surfaces is On-Demand versus a one-year commitment for the same machine, which is exactly the On-Demand vs Reserved trade-off most teams get wrong by defaulting to On-Demand and never revisiting it.

Choosing the right pricing model

There is no single cheapest model, only the cheapest model for a given workload shape. A simple decision frame that mirrors what experienced teams actually do:

• Predictable, always-on baseline: cover it with a Compute Savings Plan or Reserved Instances. Running steady workloads on On-Demand is the most common and most expensive mistake; see running predictable, cost-optimized workloads.

• Spiky or short-lived: keep it On-Demand so you pay nothing when it is off.

• Fault-tolerant and flexible: push it onto Spot for the deepest cut.

The mature pattern is not picking one model but blending all three: commitments for the floor, Spot for the elastic top, and On-Demand for the unpredictable middle. Pairing the right model with the broader set of AWS cost optimization tools is what turns a one-time pricing decision into a durable saving.

Right-size your EC2 pricing with Amnic

Choosing a model is a one-time decision; keeping it optimal as your usage drifts is continuous work. Amnic gives engineering and FinOps teams a daily, weekly, and monthly view of EC2 spend, so commitment coverage and instance choices stay matched to actual usage instead of last quarter's guess. 

Its rightsizing and commitment recommendations flag over-provisioned instances and gaps in Savings Plan coverage, while utilization tracking shows where you are paying for compute you are not using. One Amnic customer, MetaMap, used exactly this approach to lower EC2 costs by 33%.

FAQ

What is the cheapest EC2 pricing model? 

It depends on the workload. Spot is cheapest per hour (up to 90% off) but can be interrupted. For steady, interruption-sensitive workloads, Savings Plans or Reserved Instances (up to 72% off) are cheapest. On-Demand is cheapest only for short-lived or unpredictable use.

What is the difference between Reserved Instances and Savings Plans? 

Reserved Instances commit you to a specific instance type in a specific region. Savings Plans commit you to a dollar-per-hour amount of compute that applies across instance families and sizes. Savings Plans are more flexible at comparable discounts.

How much can Spot Instances actually save? 

The headline is up to 90% off On-Demand, but real blended savings typically land between 59% and 77% once interruptions and fallback capacity are factored in.

Does EC2 have a free tier? 

Yes, new accounts get 750 hours per month of t2.micro or t3.micro Linux usage for the first 12 months.

How do I estimate EC2 pricing before launching? 

Use the AWS Pricing Calculator: select region, instance type, OS, and run hours, then compare On-Demand against a one or three year commitment to see the discounted total.