AWS Lambda pricing starts at $0.20 per million requests and $0.0000166667 per GB-second of compute on x86, after a monthly free tier that never expires. The bill for the running workload actually generated in 2026 sits well above what those two rates suggest.
AWS Lambda pricing at a glance
Lambda now has three compute models, each metered differently. AWS Lambda’s pricing page splits them into Functions and MicroVMs, with Managed Instances sitting inside the Functions family as a separate billing path.
Request pricing is identical across regions and architectures. Everything else moves with region, memory, architecture, and the compute model you pick.
Every AWS account gets 1 million requests and 400,000 GB-seconds of compute a month with no expiry. That allowance stretches roughly 3.2 million seconds of execution at 128 MB but only 100,000 seconds at 4 GB, which a moderately busy function exhausts in a day. Provisioned-concurrency functions don't qualify; they bill at full rates from the first invocation.
AWS Lambda pricing breakdown
Lambda Functions on-demand pricing
What's included: Two charges. Every invocation counts as one request at $0.20 per million, whether the function succeeds, errors, or retries. Duration bills in GB-seconds, which is allocated memory multiplied by run time, rounded up to the nearest millisecond.
Best for: Event-driven workloads with variable traffic, anything that idles for long stretches, and services where request volume is hard to forecast.
Memory is configurable between 128 MB and 10,240 MB in 1 MB increments. Run time also includes initialization code declared outside your handler, so heavy runtimes with large dependency trees pay for that init on every cold start.
Aggregate monthly duration earns automatic volume discounts, introduced in August 2022 and still in effect.
Tiers apply per account and aggregate under AWS Organizations consolidated billing.
Arm functions on Graviton bill at $0.0000133334 per GB-second in the first tier against $0.0000166667 on x86, a reduction of roughly 20%. Interpreted runtimes usually move across with no code changes.
Pros:
- ✅ You pay nothing when nothing runs.
- ✅ Scaling and capacity planning disappear as operational work.
- ✅ The free allowance genuinely covers small production workloads indefinitely.
Cons:
- ❌ Unit economics degrade sharply as duration and memory climb.
- ❌ Idle wait time inside a function bills the same as active compute.
Lambda Managed Instances pricing
What's included: Three charges. Requests stay at $0.20 per million. EC2 instance charges apply for the capacity provisioned in your account. On top sits a compute management fee of 15%, calculated on the EC2 on-demand price for those instances.
Best for: Steady-state, high-volume functions, workloads needing specific hardware such as Graviton4, and anyone with unused EC2 commitment already on the books.
Managed Instances functions carry no separate per-request duration charge. Multiconcurrency lets one execution environment handle several concurrent requests, so the same traffic consumes fewer instance-hours.
Existing EC2 commitments apply to the instance charges. Compute Savings Plans and Reserved Instances reach up to 72% off EC2 on-demand.
Pros:
- ✅ Commitment discounts finally reach serverless workloads.
- ✅ Preprovisioned environments remove cold starts without a provisioned concurrency charge.
- ✅ Migration costs nothing in rewrite time, since the programming model, event sources, IAM roles, and monitoring all carry over.
Cons:
- ❌ The management fee is calculated on the on-demand instance price, so it does not shrink when your Savings Plan discount does.
- ❌ Without a commitment, the model can cost more than on-demand Lambda. The math is in the comparison below.
- ❌ Function code has to be thread-safe before it can run under multiconcurrency.
Lambda MicroVMs pricing
What's included: Per-second billing on vCPU and memory, plus separate charges for snapshot reads, snapshot writes, and snapshot storage. You configure a baseline in a 2:1 memory-to-CPU ratio, with 2 GB and 1 vCPU as the default.
Best for: Sandboxed environments for coding assistants, CI and security-scan jobs running untrusted code, and any workload needing hardware isolation per tenant.
During peak activity a MicroVM scales vertically to four times its baseline, up to 8 GB and 4 vCPU, and you pay for the additional resources only while they are active. Idle MicroVMs can be suspended for up to eight hours, with memory and disk state preserved and no compute charges while suspended.
On Graviton in us-east-1, compute runs $0.0000276944 per vCPU-second and $0.0000036667 per GB-second. Snapshot writes cost $0.0038 per GB, reads $0.00155 per GB, and image storage $0.08 per GB-month with a one-week minimum retention.
Pros:
- ✅ Vertical scaling removes the need to size every environment for its peak.
- ✅ Suspend and resume cuts the cost of idle sandboxes without discarding state.
- ✅ Snapshot-based startup removes the standby pool that latency requirements used to demand.
Cons:
- ❌ Snapshot read, write, and storage form a third cost axis to model.
- ❌ The one-week minimum retention on images makes short-lived image churn expensive.
Memory allocation and the cost-to-speed tradeoff
Memory is the single control with the largest effect on a Lambda bill, because it sets CPU allocation as well as price. Around 1,769 MB a function receives a full vCPU, and allocation scales proportionally either side of that.
A function pinned to 128 MB to save money can run so slowly that its GB-second total exceeds the same function at 512 MB. Past the sweet spot where added CPU stops shortening execution, extra memory becomes pure cost.
Which AWS Lambda pricing model should you choose?
The right billing model depends on how predictable your traffic is and whether you already hold EC2 commitments.
Choose on-demand Functions if you:
- Run traffic that is spiky, seasonal, or hard to forecast.
- Have functions that sit idle for hours at a time.
- Want zero capacity planning and are willing to pay a premium per unit of compute for it.
Choose Managed Instances if you:
- Run high, predictable request volume against a small number of functions.
- Already hold Compute Savings Plans or Reserved Instances with headroom.
- Need specific hardware, high-bandwidth networking, or Graviton4.
Choose MicroVMs if you:
- Execute untrusted or tenant-supplied code that requires hardware isolation.
- Run long-lived environments that idle between bursts of activity.
- Currently maintain a warm standby pool purely to meet startup latency targets.
3 AWS Lambda cost examples
All three use us-east-1 rates on x86 and apply the free-tier allowance every account receives.
Divide each total by its request count and the spread runs from under a dollar per million on the internal API to $32.61 per million on the event pipeline, a 36-fold difference driven entirely by duration and memory. Request charges never explain a Lambda bill; in the event pipeline they land under 1% of the total.
The charges that land outside the Lambda line item
A function invocation pulls in other services, and those services bill separately. Logging is the largest of them for a typical workload.
CloudWatch moved Lambda logs to volume tiered pricing in May 2025, dropping from $0.50 per GB to as low as $0.05 per GB as monthly volume rises. Routing logs to S3 or Data Firehose starts lower still, at $0.25 per GB.
Tiers apply per account, so a multi-account setup cannot pool its way down. Storage and query charges sit on the CloudWatch pricing page and accrue independently of ingestion.
Provisioned concurrency is the charge that runs while nothing happens. Holding 10 environments at 512 MB for a full month costs about $55 in capacity alone, before a single invocation arrives.
Our AWS cloud cost reduction playbook covers the higher-leverage cuts across CloudWatch, storage, and other services these charges roll up from.
Downstream storage costs
Every function that writes to S3, DynamoDB, or RDS produces data that somebody then retains, versions, snapshots, and protects. That spend appears on your storage and backup lines, and no Lambda cost report will connect it back to the pipeline that generated it.
63% of cloud IT leaders say storage costs force them to retain or protect less data than they should, per Eon's 2026 Cloud Data Infrastructure Report. Of that group, 87% also keep data they don't need.
Gartner's Hype Cycle for Backup and Data Protection Technologies, 2026 names Eon as a sample vendor for PaaS Backup and recommends pricing PaaS-service backup alongside the cost of hosting the application itself. PaaS Backup currently sits in the report's "Sliding into the Trough" phase.
Eon's Cost Explorer applies that model on the storage side: resources are classified as they are created, retention runs on policy, and backup spend is visible per resource across every account.
Cloud teams working this way cut backup storage costs 30–50%; StructuredWeb landed at 40% in storage and restoration savings. If your Lambda functions feed a data platform, that spend belongs in the same budget as the invocation cost.
Is AWS Lambda worth the cost?
Lambda is worth it when your traffic pattern goes quiet for real stretches, because paying nothing through those idle windows outweighs a higher unit rate during the busy periods.
It stops being worth it when compute runs continuously, when functions hold memory while waiting on something external, or when a single workload dominates your invocation count.
Lambda is worth it if you:
- Have genuine idle periods where scale-to-zero pays for itself.
- Run event-driven glue code where operational overhead costs more than compute.
- Need to ship without provisioning anything.
Skip Lambda if you:
- Run steady, continuous compute where a committed instance is cheaper per unit.
- Have functions that spend their duration waiting on external calls.
- Need runtimes, execution times, or hardware Lambda does not support.
AWS Lambda alternatives and pricing comparison
A single Fargate task at 1 vCPU and 2 GB running around the clock costs about $36 a month. Ten million Lambda invocations at 1 GB and 200 ms cost roughly the same. Which side wins depends on how much concurrency one Fargate task actually absorbs in your workload.
EC2 sits further along the same line. An m7g.xlarge runs about $119 a month on demand and drops sharply under a Savings Plan or Reserved Instance. It wins on unit cost for anything that runs continuously, and it hands you back patching, scaling, and capacity planning as work your engineers own.
AWS prices a 100-million-request service at 200 ms on m7g.xlarge with a three-year Compute Savings Plan at $160.36 a month, roughly 54% less than the $346 the same profile would cost on on-demand Lambda.
Without that commitment, those instance-hours at on-demand EC2 rates land past $395, more than on-demand Lambda charges. Price Managed Instances against your existing EC2 coverage before you migrate anything.
Compute Savings Plans cover Lambda duration and provisioned concurrency, though not request charges. If you already hold one for EC2 or Fargate, the commitment applies across all three.
6 ways to lower AWS Lambda costs
Six moves cover the highest-leverage cost cuts on Lambda, roughly in order of impact:
1. Move eligible functions to Arm. Graviton carries a lower GB-second rate at the same or better performance, and it reaches its volume tiers later than x86. Interpreted runtimes usually need no code changes, making this the cleanest reduction available.
2. Right-size memory against measured duration. Test each function across memory settings and plot cost against runtime. Lambda Power Tuning runs this as a Step Functions workflow and returns the optimum. Retest after significant code changes.
3. Cut log volume at the source. Drop debug logging in production, set explicit retention on every log group, and route high-volume logs to S3 or Data Firehose. Log groups default to never expiring, which turns ingestion into a permanent storage charge.
4. Remove billable wait time. A function polling an external API bills for every millisecond it spends waiting. Move the wait into Step Functions or durable functions, where execution suspends and duration charges stop until work resumes.
5. Move steady functions to Managed Instances. Identify functions with flat, predictable load and high invocation counts. Those are the ones where committed EC2 pricing beats per-request duration billing, provided you hold the commitment.
6. Audit provisioned concurrency schedules. Capacity configured for a launch or a sale often stays enabled afterward. Provisioned concurrency bills from enable to disable, rounded to the nearest five minutes, whether traffic arrives or not.
For modeling before you build, the AWS Pricing Calculator handles requests, duration, and the common add-ons. Our AWS cost optimization tools comparison covers the platforms that do continuous attribution once the workload is live.
Where to focus your AWS Lambda pricing work
AWS Lambda pricing rewards attention in this order: memory first, architecture second, then logging, and the compute model last. Invocation rate almost never decides the bill.
For the storage, retention, and protection that pile up outside the Lambda line item, our breakdowns of AWS Backup pricing, S3 lifecycle and versioning costs, and cloud cost optimization walk through the math.
Can you see what your serverless pipelines cost after the invocation ends? Book a demo and see how Eon's Cost Explorer surfaces backup spend per resource across every account, with classification and retention automated behind it.
Frequently asked questions
How much does AWS Lambda cost per million requests?
AWS Lambda charges $0.20 per million requests, and that rate is the same in every region and on both architectures. Duration charges come on top and usually dominate. Across three published AWS examples, the all-in cost per million requests ranged from $0.91 to $33.53 depending on memory and execution time.
Is AWS Lambda free?
Yes, within limits. Every AWS account receives 1 million requests and 400,000 GB-seconds of compute each month, and that allowance does not expire. Accounts created on or after July 15, 2025 sit on a restructured free tier with $100 to $200 in credits, though the always-free Lambda allowance applies to them the same way.
What is a GB-second in AWS Lambda pricing?
A GB-second is one gigabyte of memory held for one second of execution. Multiply allocated memory in gigabytes by run time in seconds by invocation count to get your monthly total. Duration rounds up to the nearest millisecond, and the calculation includes initialization code outside your handler.
Does Arm (Graviton) lower AWS Lambda costs?
Yes. Arm functions bill at $0.0000133334 per GB-second in the first tier against $0.0000166667 on x86, roughly 20% less, with equal or better performance across common workloads. Arm also reaches its tiered discounts at higher volume thresholds than x86.
When does AWS Lambda cost more than EC2 or Fargate?
Lambda costs more once compute runs continuously at high memory. The crossover depends on how much concurrency one container or instance absorbs, so model it against your own throughput. Duration is the useful signal. Functions measured in seconds cross over far earlier than functions measured in milliseconds.
Do Compute Savings Plans cover AWS Lambda?
Yes, for duration and provisioned concurrency, though not for request charges. The same commitment also applies to EC2 and Fargate, so unused coverage moves between compute services. Lambda Managed Instances draw on EC2 pricing options directly, including Reserved Instances.
