What is the difference between Savings Plans and Reserved Instances?

Both are AWS billing commitments that trade a 1- or 3-year term for a lower rate than On-Demand. They differ in what you commit to.

A Savings Plan commits you to an hourly spend amount, such as $10 per hour of compute usage at Savings Plan rates. AWS applies the discounted rate to eligible usage until that hourly amount is used, and anything above it bills at On-Demand. A Compute Savings Plan applies across EC2 instance families, sizes, Regions, and operating systems, and also covers Fargate and Lambda. An EC2 Instance Savings Plan narrows this to one instance family in one Region in exchange for a deeper discount.

A Reserved Instance commits you to a specific instance configuration, defined by attributes such as instance type or family, Region, platform, and tenancy. The discount applies only to running usage that matches those attributes.

Teams buy either one for the same reason. Some compute runs every hour regardless of traffic, such as baseline API fleets, databases, and always-on workers. Paying On-Demand rates for capacity you know you will use wastes money.

The difference in what you commit to determines how each behaves when your architecture changes. Suppose you move a service from x86 to Graviton, from EC2 to containers on Fargate, or to another Region. A Compute Savings Plan follows the spend to the new compute without any action from you. A Reserved Instance stays tied to the old configuration. If nothing still matches its attributes, you keep paying for a discount that applies to nothing, and you pay On-Demand for the new compute.

How flexible is each option, and what does flexibility cost?

That trade-off sits on a spectrum. The more of your future configuration you lock in today, the larger the discount AWS generally offers, because AWS carries less risk that you'll shift to capacity it didn't plan for.

Option What can change without penalty Exit routes
Compute Savings Plan Instance family, size, region, OS, tenancy; also covers Fargate and Lambda None. You pay the commitment for the full term
EC2 Instance Savings Plan Size, OS, tenancy, AZ within one family in one region None
Convertible RI Family, OS, tenancy, via exchange Exchange for equal or greater value, sometimes with a true-up payment
Standard RI Limited modifications such as AZ or size within a family Resale on the RI Marketplace, if eligible

Neither exit is free. A Convertible RI exchange can't reduce your total commitment, so it only reshapes spend. A Standard RI sale depends on a buyer, a price you may have to discount, and a service fee. Savings Plans have no exit at all.

To compare options, use effective discount, not the headline rate:

Effective discount = 1 − (1 − d) / u, where d is the list discount and u is expected utilization.

Illustrative estimate (assumed numbers, not a benchmark): a rigid option at d = 40% with u = 80% gives 1 − 0.60/0.80 = 25%. A flexible option at d = 30% with u = 98% gives 1 − 0.70/0.98 ≈ 28.6%. In this case the smaller headline discount saves more.

Rates and terms vary by region, payment option, and term length, and they change over time. Before you plug in d, pull the current figures from the purchase recommendations and pricing pages in your own account.

How are commitment discounts applied each hour?

Billing evaluates commitments one hour at a time. Each hour, eligible usage is matched against what you committed to, and anything that doesn't fit falls back to on-demand rates.

For a Savings Plan, the commitment is a dollar amount per hour, measured at the discounted Savings Plan rates. Eligible usage in that hour is charged at those rates and draws down the commitment. When the commitment is used up, the rest of the usage is billed at on-demand prices. If the hour ends with commitment left over, you still pay for the remainder. It does not roll forward to the next hour. Reserved Instances work the same way, except they match on instance attributes rather than a dollar pool.

Illustrative estimate. Assume a $10/hour commitment:

  • Busy hour. Eligible usage would cost $14 at Savings Plan rates. The first $10 is covered, and the remaining usage is billed at on-demand rates, which are higher than the $4 it would have cost under the plan.
  • Quiet hour. Usage comes to $6. You still pay $10, and $4 is waste.

The commitment is a floor on spend, not a cap. Each hour you either use it or lose it. A nightly scale-down to near zero means paying for idle commitment every night.

Consolidated billing complicates this. By default, discounts can apply to matching usage in any account in the organization, not only the account that bought the commitment. A platform team's plan may end up discounting another team's batch jobs. The buyer's cost report then shows unused commitment while the other team gets cheaper usage it never paid for. Showback needs explicit allocation rules. Otherwise per-team cost numbers will mislead whoever reads them.

 Consolidated billing org: one billing hour
 +--------------+  +--------------+  +--------------+
 | Platform     |  | Team B       |  | Team C       |
 | (buys plan)  |  | batch jobs   |  | services     |
 +------+-------+  +------+-------+  +------+-------+
        |                 |                 |
        +-----------------+-----------------+
                          | eligible usage
                          v
           +-----------------------------+
           | Hourly match vs commitment  |
           | SP: $10/hr at SP rates      |
           | RI: match on attributes     |
           +--------------+--------------+
                          |
        +-----------------+-----------------+
        v                 v                 v
 +--------------+  +--------------+  +--------------+
 | Covered      |  | Overflow     |  | Unused       |
 | drawn down   |  | billed at    |  | still billed |
 | at SP rates  |  | on-demand    |  | no rollover  |
 +--------------+  +--------------+  +--------------+
 Busy:  $14 at SP rates -> $10 covered, rest billed at on-demand
 Quiet: $6 used -> billed $10, $4 wasted. A floor, not a cap.
 +------------------------------------------------------------+
 | Showback: plan may discount Team B usage, while the buyer  |
 | reports unused commitment. Set explicit allocation rules.  |
 +------------------------------------------------------------+

Each hour, eligible usage from any account in the organization draws down the commitment. Usage beyond it is billed at on-demand rates, and any unused balance is still billed, using illustrative $10/hour figures.

How do you size a Compute Savings Plan from usage history?

Export at least four to eight recent weeks of hourly cost data and filter it to usage the plan can cover: eligible compute, priced at on-demand-equivalent rates, with credits and refunds removed. Sum the cost per hour. Then take a low percentile of those hourly totals, such as the 5th or 10th, instead of the mean. The mean includes peaks that autoscaling will remove, so a commitment sized to it will sit unused overnight and on weekends. A low percentile approximates the floor you pay for in nearly every hour. The query example that follows groups exported rows by hour and applies a percentile function to the totals. Commit to that floor and leave the variable portion on on-demand or spot.

Illustrative estimate (assumed numbers, not a benchmark)

  • Assumed baseline: $100/hr on-demand-equivalent spend.
  • Assumed discount: 30%.
  • Assumed over-sized plan: sized to cover $130/hr.
  • Assumed post-migration usage: $60/hr.

The formula is: hourly commitment = covered on-demand spend × (1 − discount). The plan costs more than staying on-demand whenever usage falls below the commitment, so the break-even usage equals the commitment. In this example the over-sized plan commits $91/hr, which means any hour with less than $91 of on-demand usage loses money.

Scenario Hourly commitment Usage (on-demand equiv.) Paid with plan Paid on-demand only
Right-sized, before migration $70 $100 $70 $100
Right-sized, after migration $70 $60 $70 $60
Over-sized, after migration $91 $60 $91 $60

A right-sized commitment still loses $10/hr after the migration, but that is about a third of the over-sized plan's $31/hr loss.

# Size a Compute Savings Plan from a low percentile of hourly on-demand-equivalent spend.
# Usage: python3 -I size_plan.py cost_export.csv
# Column names follow the legacy CUR CSV layout; rename them to match your export.
import sys
import pandas as pd

TIME_COL = "lineItem/UsageStartDate"
TYPE_COL = "lineItem/LineItemType"
PRODUCT_COL = "lineItem/ProductCode"
OD_COST_COL = "pricing/publicOnDemandCost"  # on-demand-equivalent price, not what you paid

# Keep usage rows only. This drops credits, refunds, taxes and fees.
# Include rows already covered by plans so the result is total demand, not just the uncovered remainder.
USAGE_TYPES = {"Usage", "SavingsPlanCoveredUsage", "DiscountedUsage"}
# Coarse filter. Narrow it further, for example by excluding EBS and data transfer under AmazonEC2.
PRODUCTS = {"AmazonEC2", "AmazonECS", "AWSLambda"}

df = pd.read_csv(sys.argv[1], usecols=[TIME_COL, TYPE_COL, PRODUCT_COL, OD_COST_COL])
df = df[df[TYPE_COL].isin(USAGE_TYPES) & df[PRODUCT_COL].isin(PRODUCTS)]

# Sum spend per clock hour.
ts = pd.to_datetime(df[TIME_COL], utc=True).dt.floor("h")
hourly = df[OD_COST_COL].groupby(ts).sum()

# Fill hours that have no rows with 0. A missing hour is a real low and must count toward the floor.
full_range = pd.date_range(hourly.index.min(), hourly.index.max(), freq="h")
hourly = hourly.reindex(full_range, fill_value=0.0)

ASSUMED_DISCOUNT = 0.30  # assumed; use the rate quoted for your term and payment option
p05, p10 = hourly.quantile(0.05), hourly.quantile(0.10)
print(f"hours={len(hourly)} mean=${hourly.mean():.2f} p05=${p05:.2f} p10=${p10:.2f}")
# The commitment is a $/hr amount at plan rates, so convert the on-demand floor.
print(f"suggested commitment ~ ${p05 * (1 - ASSUMED_DISCOUNT):.2f}/hr (sized to the p05 floor)")

A pandas script that filters an exported cost CSV to plan-eligible compute usage, sums on-demand-equivalent spend per hour with empty hours counted as zero, and converts the 5th-percentile floor into an hourly commitment using an assumed discount.

Why should you prioritize utilization over commitment coverage?

Two metrics describe a commitment, and they measure different things. Commitment coverage is the share of your eligible usage that receives the discounted rate. Utilization is the share of the commitment you actually consume. If you commit to $60/hr and your eligible usage that hour is $80, utilization is 100% and coverage is 75%. If usage drops to $45, coverage is 100% but utilization falls to 75%. You still pay for the idle $15.

The two metrics work against each other because usage fluctuates. The only way to reach 100% coverage is to commit at or near your peak, so every hour below that peak burns unused commitment. Your true baseline is the floor that usage rarely drops below. Any commitment above that floor is a bet that the floor will rise. Uncovered usage costs you only the difference between the discounted and on-demand rate. Unused commitment costs you the full committed rate. That asymmetry is why utilization should come first.

Aim for utilization close to 100% and accept coverage well below that. Some on-demand spend is not waste. It is the price of being able to scale down.

The same reasoning applies to term length and payment option. A three-year, all-upfront purchase gets the best rate, but it ties up cash and fixes your assumptions for longer. Over three years, the instance families, regions and runtimes you depend on may change, while the commitment stays the same.

Which engineering changes should trigger a commitment review?

Each of the changes below should trigger a commitment review before it ships. The risk is over-commitment: paying every hour for capacity the new architecture no longer uses. The instrument you hold determines whether a migration also becomes a billing project.

Engineering change Standard RI Convertible RI EC2 Instance Savings Plan Compute Savings Plan
Move to ARM-based instances (new family) Not covered; resell or absorb Exchange required for an ARM family Not covered; bound to one family Covered automatically
Containerize onto EC2 hosts Covered only if hosts match the reserved family Covered after an exchange if the family changes Covered if the family is unchanged Covered
Shift work to Fargate or Lambda Not covered Not covered Not covered Covered
Consolidate regions Not covered Not covered; exchanges stay in-region Not covered Covered
Rightsize within a family Size flexibility applies to regional Linux RIs Same as Standard Covered across sizes Covered

Rightsizing deserves extra care. Every instrument handles a size change. None of them handles a smaller total. If rightsizing cuts hourly spend below your commitment, the gap is waste whatever discount you hold.

The flexibility also affects engineering time. A team holding Standard RIs on x86 instances has to plan an ARM migration around remaining terms, resale options or exchange paperwork. That is time spent on billing instead of on the migration itself.

With a Compute Savings Plan, the team migrates and the discount follows the spend. The higher rate compared with a Standard RI is the price of removing that coordination step.

When do Reserved Instances still win, and when should you avoid committing at all?

Standard RIs still make sense in a few specific cases where the extra discount outweighs the lock-in.

Where Reserved Instances still make sense

  • Narrow, fixed fleets. Some workloads run the same instance family in the same region for years, such as a stateful queue cluster or a licensed application pinned to specific hardware. The flexibility of a Savings Plan does nothing for these, so the deeper Standard RI discount is the better deal. The lock-in is only a risk if something changes, and on these fleets nothing does.
  • Capacity guarantees. Savings Plans are purely a billing construct and do not reserve capacity. A zonal RI does reserve capacity in a specific Availability Zone. This matters if you need to be sure instances can launch during a regional surge or a failover. Another option is to pair a Savings Plan with a separate capacity reservation, but compare the cost of both approaches.
  • Services outside EC2. Managed databases, caches and search clusters have historically used their own reserved-node or reserved-instance models, and Savings Plans have not always covered them. Coverage rules change, so check current eligibility in your own account's billing console rather than assuming.

When not to commit at all

A commitment assumes a steady baseline exists. If usage spikes unpredictably or is trending down, any commitment you size today will go underused within months. For these workloads, work on the variable cost first:

  • Move interruptible work to Spot.
  • Tighten autoscaling so idle capacity drains.
  • Rightsize oversized instances.

Each of these lowers the baseline. Commit only after it has stopped moving.

How should teams layer purchases and own commitment decisions?

Buy in tranches, not in one large purchase

Avoid committing to the full target in a single purchase. Split it into smaller tranches bought at regular intervals, for example one portion each quarter. Each tranche then expires on a different date, so you never have to renew your entire commitment at once against a single snapshot of usage. A mistake also costs less. If a migration or a traffic change shrinks your baseline, only the most recent tranche is oversized, and the older ones expire on their normal schedule. Each expiration is a scheduled point to re-check the baseline before you renew.

Centralize ownership

In multi-team or multi-account organizations, commitments should belong to one function, usually a central platform or FinOps team. Commitment discounts are typically shared across accounts in an organization, so a team that buys on its own may end up subsidizing usage elsewhere or overlapping with another team's purchase. The owning team works from a fixed review cadence, tracks utilization and coverage, and publishes showback rules that state how discounts are allocated back to each team's costs.

Why this helps engineers

Product teams no longer need to react to cost alerts or get pulled into ad hoc purchase debates. They report planned changes, such as a move to ARM-based instances, containerization, or a region shift, through the review process. The central team adjusts upcoming tranches to account for them. Because no single team carries a commitment of its own, teams can change their architecture without first checking whether it would strand a reservation.

Alternatives to consider

Commitments are one way to lower the unit price of compute. Several other approaches attack the same bill from different angles, and most of them should come before or alongside a commitment.

Reduce usage before committing to it. Rightsizing over-provisioned instances, deleting idle environments, and scheduling non-production fleets to shut down outside working hours all shrink the baseline itself. Doing this first matters because a commitment sized to a wasteful baseline locks that waste in at a discount, and that is still waste.

Spot Instances for interruptible work. Spot capacity is often discounted more heavily than any commitment and needs no term. The trade-off is that AWS can reclaim instances with a two-minute warning. That suits stateless workers, CI runners, batch jobs, and queue consumers that checkpoint progress. It is a poor fit for stateful services or anything with strict latency requirements. Spot and Savings Plans work well together: commit to the steady floor, and serve elastic or fault-tolerant load with Spot.

On-Demand Capacity Reservations. These guarantee capacity in a specific Availability Zone but provide no discount by themselves. Use them when the risk you need to manage is running out of capacity during a launch or failover, not the price. A Savings Plan can apply to the reserved usage to lower its cost.

Architecture changes. Moving to cheaper instance families or processor architectures, or moving spiky workloads to request-based services, lowers the price of each unit of compute. Make these moves before a large purchase so the commitment matches the cheaper baseline.

How OneEnv can help

Disclosure: OneEnv is the product behind this blog.

OneEnv (oneenv.io) is a workspace-based governance platform for the things services share — environment variables, database structure, API routes and validation rules. A change to a shared value is staged on a branch, reviewed once, and then every affected service explicitly approves or rejects it before it goes live, so you can see who a change touches before it ships instead of finding out in production.

Applied to savings plans vs reserved instances, this is where it helps:

Saves time

  • Every service that inherits a shared value is listed in one place, so nobody searches repos and chats to find out who is affected.
  • One review per branch replaces coordinating each team separately; each service still keeps its own approve/reject.
  • Changing a shared variable is one edit and one review instead of editing every service's .env by hand.

Saves cost

  • Fewer production incidents and rollbacks caused by a shared value, route or column changing under a consumer unannounced.
  • Less duplicated configuration to maintain: one shared definition inherited by many services.
  • A wrong or stale value is caught at review time by the services it affects, not by a failed deploy.
  • OneEnv has a free plan (1 workspace, unlimited services and members, full governance), so trying it costs nothing.

These are qualitative benefits, not measured results — how much you save depends on how many services share the value.

Want to try it? Start at oneenv.io.