AWS Pricing Changes in 2026: Every Update That Affects Your Bill

In 2026, the two most impactful AWS pricing shifts for most teams are: the public IPv4 charge ($3.65/month per IP, in effect since February 2024) and Graviton4 general availability, which delivers up to 30% better price-performance than Graviton3 (AWS, 2026). AWS pricing changes fall into two categories — reductions tied to new hardware generations and increases tied to charges for previously free resources. Both are in play simultaneously. For a side-by-side view of where AWS versus Azure undercuts each other, see our AWS vs Azure pricing comparison.

The single largest unresolved cost impact for most teams is the public IPv4 charge introduced February 1, 2024: $0.005/hour per public IP, or $3.65/month. It applies to every public IPv4 attached to an EC2 instance, RDS instance, load balancer, NAT Gateway, or any other resource — including ones in active use (AWS Blog, Feb 2024). Many teams that built infrastructure before 2024 have never audited their IPv4 footprint. That audit is free and takes under an hour.

On the savings side, Graviton4 instances (r8g, c8g, m8g families) reached full general availability in early 2026. AWS claims up to 30% better compute performance per dollar compared to Graviton3, with memory-optimized r8g seeing the largest gains. Any team on r6g or r7g instances should evaluate migration — most Linux workloads require zero code changes.

Three confirmed pricing changes took effect in Q1 2026. Additional changes expected through Q4 are detailed below. Check back or subscribe for updates as AWS announces them — AWS typically provides 30 days advance notice for increases; reductions take effect immediately.

Q1 2026 (January – March)

Q2–Q4 2026 (Anticipated)

AWS historically announces pricing changes at re:Invent (December) with rollouts through Q1–Q2 of the following year. Additional changes expected for the remainder of 2026:

• EC2 Savings Plans rate refreshes: Expect updated 1-year and 3-year rates for Graviton4 families as adoption scales and AWS adjusts competitive positioning.
• S3 Intelligent-Tiering threshold changes: AWS has periodically adjusted the minimum object size for automatic tiering (128 KB as of Q1 2026). A raise could affect small-object workloads.
• Aurora Serverless v2 ACU pricing: Competitive pressure from PlanetScale, Neon, and Turso is pushing AWS to sharpen serverless database pricing. A reduction in ACU pricing or minimum ACU floor is plausible in H2 2026.
• AWS Lambda duration pricing: Lambda pricing (per GB-second) has remained stable since 2018. A reduction would be significant for serverless-heavy architectures.

Three 2024–2025 pricing changes are not yet reflected in most teams' cost models. If your infrastructure predates them and you have not audited recently, these are almost certainly hitting your bill without anyone noticing. The IPv4 charge alone adds over $1,100/year for a startup with a typical 27-IP footprint — a recurring cost invisible to most billing dashboards until you filter for it explicitly.

Compute is 40–60% of a typical AWS bill and carries more pricing levers than any other category (AWS Cost Explorer benchmarks, 2025). In 2026, the dominant story is Graviton4 broad availability paired with Savings Plans. Teams migrating from x86 on-demand to Graviton4 Savings Plans can reduce compute costs by 50–65% for equivalent workloads. That is not a marginal improvement — it is a structural one.

EC2 On-Demand Pricing Trends

AWS reduces on-demand prices for older instance families when newer generations arrive. In 2026, the m5/c5/r5 families (Intel Skylake/Cascade Lake) have seen 3–7% reductions in select regions, following the same pattern seen when m7i launched and m6i prices were trimmed. On-demand pricing for current-generation x86 (m7i, c7i, r7i) remains stable.

Graviton4 (r8g, c8g, m8g) is priced approximately 10–15% lower than equivalent Graviton3 instances on-demand, while delivering 25–30% more performance per core. On a price-performance basis, Graviton4 is the best value compute option AWS currently offers for general Linux workloads.

Graviton Discounts and Savings Plans

Compute Savings Plans for Graviton4 families offer 1-year and 3-year terms. One-year no-upfront plans deliver approximately 30% off on-demand rates. Three-year all-upfront plans reach 50–55% off. Combined with Graviton4's baseline efficiency improvement, teams migrating from x86 on-demand to Graviton4 Savings Plans can reduce compute costs by 50–65% for equivalent workloads (AWS Savings Plans Pricing). For a full breakdown of which commitment model fits your workload, see our reserved vs spot vs on-demand comparison.

Spot Instance Market Conditions in 2026

Spot prices in 2026 remain volatile in us-east-1 for GPU instances (p4d, p5, g5) due to sustained AI/ML training demand. General-purpose Spot (m7g, c7g) has remained relatively stable at 60–70% discounts from on-demand. For batch, CI/CD, and stateless workloads, Graviton-based Spot instances (m8g, c8g) offer particularly attractive interruption rates in many availability zones due to lower overall demand pressure compared to equivalent x86 families.

Every gp2 EBS volume in your account is a guaranteed 20% cost reduction waiting to happen — gp3 costs $0.08/GB-month versus $0.10/GB-month for gp2, at better baseline performance, with zero downtime to migrate (AWS EBS Pricing, 2026). Storage is the most underestimated optimization target: the changes are low-risk, the savings are immediate, and in 2026 most AWS accounts still have unmigrated gp2 volumes.

S3 Express One Zone — When It's Worth It

Amazon S3 Express One Zone became generally available in late 2023 and has matured through 2025–2026. It delivers single-digit millisecond latency and up to 10x higher request throughput than standard S3. The pricing model reflects the premium: $0.16/GB-month for storage (versus $0.023/GB-month for S3 Standard) and $0.0025/ $0.0008 per 1,000 PUT/GET requests (AWS S3 Pricing).

Express One Zone is not a replacement for standard S3. It is purpose-built for latency-sensitive workloads: ML training data access, high-frequency analytics buffers, and applications where sub-10ms S3 response time materially affects throughput. For everything else, standard S3 remains substantially cheaper.

EBS gp3 vs. gp2 — Still Worth Migrating in 2026

EBS gp3 was introduced in 2020. It is 20% cheaper than gp2 ($0.08/GB-month vs. $0.10/GB-month). Its baseline performance (3,000 IOPS, 125 MB/s throughput) exceeds gp2's baseline for volumes under 1 TB at no extra cost (AWS EBS Pricing).

Migration is live — no downtime required. Change volume type directly in the EC2 console or via CLI. Filter for "gp2" in the EC2 Volumes console right now. Every gp2 volume is money being lost.

S3 Intelligent-Tiering — Thresholds to Know

S3 Intelligent-Tiering automatically moves objects between access tiers based on access patterns. In 2026, the minimum object size that qualifies for automatic tiering is 128 KB — objects smaller than this are stored in Frequent Access tier regardless of access patterns. For workloads with many small objects (under 128 KB), the monitoring charge ($0.0025 per 1,000 objects/month) can exceed any tiering savings. Profile your object size distribution before enabling Intelligent-Tiering at scale.

Glacier Instant Retrieval vs. Flexible Retrieval

Glacier Instant Retrieval ($0.004/GB-month) and Glacier Flexible Retrieval ($0.0036/GB-month) serve different access patterns. In 2026, many teams default to Glacier Instant Retrieval for compliance archiving when Flexible Retrieval would cost less for data accessed rarely or never. Flexible Retrieval bulk restore (5–12 hours) costs $0.00/GB — only pay for compute. If you're archiving logs, choose the tier that matches your actual recovery time objective.

Networking is the most opaque AWS pricing category and the source of the largest billing surprises. The true cost of internet-bound traffic through a NAT Gateway is $0.135/GB — three times what most teams assume when they see only the $0.045 processing fee on the pricing page (AWS VPC Pricing, 2026). In 2026, the three compounding networking charges are egress, NAT Gateway processing, and the IPv4 charge affecting every public resource.

Data Egress: Still $0.09/GB (First 10 TB)

AWS data transfer out to the internet remains $0.09/GB for the first 10 TB/month in us-east-1, dropping to $0.085/GB from 10–50 TB and $0.07/GB beyond 50 TB. AWS has not reduced egress pricing in response to competitive pressure despite Google Cloud's free egress policy for services like Cloud Run. Teams with significant egress (video delivery, large file downloads, API responses with large payloads) should consider CloudFront — CloudFront origin fetches from AWS are free, and CloudFront's own egress pricing is lower than direct S3/EC2 egress for most regional pricing tiers (AWS CloudFront Pricing).

NAT Gateway: The True Cost Is $0.135/GB

NAT Gateway pricing in 2026 is unchanged: $0.045/hour per gateway ($32.40/month) plus $0.045/GB data processing. But the real cost of internet-bound traffic through a NAT Gateway is $0.135/GB — processing ($0.045) plus standard egress ($0.09). That is three times what most teams assume when they see only the "processing" line on the pricing page (AWS VPC Pricing). For a deep dive on reducing this cost, see our AWS NAT Gateway pricing guide.

Additionally, cross-AZ NAT traffic adds $0.01/GB each way. A three-AZ deployment with a single centralized NAT Gateway in one AZ means two-thirds of traffic incurs this cross-AZ penalty on top of the base rates.

The single highest-impact fix: add VPC Gateway Endpoints for S3 and DynamoDB. They are free and take five minutes to configure. All S3 and DynamoDB traffic from private subnets then bypasses NAT entirely — no processing fee, no egress fee. Most teams find this eliminates 30–50% of their NAT Gateway traffic immediately.

IPv4 Charges — Action Required

At $3.65/month per public IP, IPv4 charges have been accumulating since February 2024. Many organizations that built their infrastructure before this date have never audited their public IP footprint. Use the Public IP Insights dashboard (VPC console → "Public IP Insights") or AWS Config rule "ec2-no-public-ip" to surface all public IPs. Prioritize:

• RDS instances with public accessibility enabled but accessed only from within the VPC
• EC2 instances in private subnets that were accidentally assigned public IPs
• Old Elastic IPs associated with stopped or terminated instances
• Internal load balancers (ALB/NLB) that were deployed as internet-facing by default

Aurora Serverless v2 carries a minimum cost of $43.20/month per cluster even at zero traffic — because the 0.5 ACU floor ($0.06/hour) never goes to zero (AWS Aurora Pricing, 2026). Teams running separate Aurora Serverless v2 clusters for dev, staging, and prod pay $130+/month in minimum standby charges before a single query runs. For dev environments, provisioned db.t4g instances under $30/month are often the cheaper choice.

RDS Pricing in 2026

RDS on-demand instance pricing follows the same generation pattern as EC2. In 2026, db.r8g and db.m8g instance families (Graviton4) are available for RDS and deliver the same 25–30% price-performance improvement over Graviton3 equivalents. Teams still running db.r5 or db.m5 instances should evaluate a migration path.

RDS Multi-AZ pricing is approximately 2x single-AZ. For staging and development environments, single-AZ RDS with daily automated snapshots can reduce database infrastructure costs by 40–50% for non-production workloads.

One often-missed charge: RDS storage autoscaling. Enabling it with a high maximum threshold lets AWS silently scale your storage up during a traffic spike — and storage is never automatically reduced when traffic normalizes. Monitor allocated RDS storage in Cost Explorer filtered by "RDS Provisioned Storage."

Aurora Serverless v2 — ACU Billing Nuances

Aurora Serverless v2 is billed per Aurora Capacity Unit (ACU) per hour. In us-east-1, ACU pricing is $0.12/ACU-hour for PostgreSQL-compatible and $0.12/ACU-hour for MySQL-compatible. The minimum ACU setting defaults to 0.5 ACU, which means even at zero traffic, Aurora Serverless v2 costs at minimum $0.06/hour ($43.20/month) per cluster — just for being available (AWS Aurora Pricing).

Teams running multiple Aurora Serverless v2 clusters for separate environments (dev, staging, prod) will pay $43+ per month per cluster in minimum standby costs. For development environments that are idle most of the time, provisioned db.t4g instances (<$30/month for db.t4g.medium) often cost less than the Serverless v2 minimum floor.

Monitor ACU utilization and set CloudWatch alarms on "ServerlessDatabaseCapacity" to detect runaway scaling. The Serverless v2 maximum ACU setting is a hard cap — configure it explicitly to prevent unexpected scale-up costs.

RDS Public IPv4 Charges

RDS instances with "Publicly accessible" enabled are subject to the $3.65/month IPv4 charge introduced in February 2024. Many RDS instances have public accessibility turned on by default from initial deployment, even when they are only accessed from within the VPC via private subnets. Disabling public accessibility on RDS instances that do not need it removes the IPv4 charge entirely and reduces attack surface simultaneously.

EKS charges $0.10/hour per cluster — $72/month — regardless of how many nodes or workloads run inside it (AWS EKS Pricing, 2026). This flat fee has been in place since 2020 but continues to catch teams off-guard. A team running 5 clusters for environment separation pays $360/month in cluster fees before a single EC2 node is billed. Organizations using "one cluster per team" patterns can reach $720–$1,440/month in cluster fees alone.

EKS Cluster Fee Impact by Architecture

A team running 5 clusters (prod, staging, dev, canary, shared-services) pays $360/month in cluster fees before a single EC2 node is spun up. Organizations that adopted "one cluster per team" patterns can find themselves with 10–20 clusters and $720–$1,440/month in flat cluster fees.

In 2026, the trend is toward cluster consolidation — using namespaces, NetworkPolicies, and RBAC to isolate environments within fewer clusters rather than provisioning a cluster per environment. EKS managed node groups on Graviton4 (m8g, c8g) provide the best cost efficiency for general Kubernetes workloads.

Fargate for EKS — When It Costs More

EKS with Fargate (serverless pods) is priced at $0.04048/vCPU-hour and $0.004445/GB-hour. For always-on workloads, Fargate is more expensive than equivalent EC2 nodes. Fargate makes economic sense for workloads that are genuinely bursty and unpredictable, where the alternative would be significant over-provisioning on EC2. For steady-state workloads, EC2 node groups with Graviton4 Spot instances are substantially cheaper.

EKS Auto Mode (2025–2026)

AWS launched EKS Auto Mode in late 2024, which automates node provisioning, scaling, and lifecycle management. In 2026, EKS Auto Mode carries an additional compute surcharge (approximately 2% above EC2 on-demand for managed nodes) in exchange for fully automated node management. For teams without dedicated platform engineering capacity, this trade-off may be worthwhile — but teams with mature Karpenter deployments likely get more control and lower cost without Auto Mode.

Bedrock Knowledge Bases (managed RAG) introduced an often-missed cost floor in 2025: OpenSearch Serverless as the default vector store carries a minimum charge of $172/month even when idle — 2 OCUs at $0.24/hour (AWS OpenSearch Pricing, 2026). Teams building RAG applications need to factor this into their cost model before launch, or evaluate pgvector on RDS as a zero-minimum alternative.

Amazon Bedrock Pricing in 2026

Bedrock is the primary entry point for teams consuming foundation model APIs on AWS. On-demand pricing is per token (input and output billed separately). Key rates as of March 2026 (AWS Bedrock Pricing):

• Claude Haiku (4.5): ~$0.0008/1K input tokens, ~$0.004/1K output tokens (reduced from prior generation)
• Claude Sonnet (4.6): ~$0.003/1K input, ~$0.015/1K output
• Meta Llama (8B): ~$0.0003/1K input, ~$0.0006/1K output (reduced ~15%)
• Amazon Titan Text Express: ~$0.0002/1K input, ~$0.0006/1K output

For high-volume inference (over $500/month on Bedrock), provisioned throughput is often more cost-effective than on-demand. Provisioned throughput requires a 1-month or 6-month commitment but provides dedicated capacity and a lower effective per-token rate.

SageMaker Pricing Changes

SageMaker pricing continues to evolve as AWS adds managed inference capabilities. Key changes affecting 2026 bills:

• SageMaker Inference: ml.inf2 instances (AWS Inferentia2) continue to offer 40–60% lower inference costs vs. GPU instances for supported model architectures. If you're serving transformer models on SageMaker using ml.g5 instances, benchmark ml.inf2 — the cost difference at scale is significant.
• SageMaker Studio: Idle SageMaker Studio kernels accrue hourly instance charges. A team with 10 data scientists who frequently leave ml.t3.medium kernels running overnight pays an avoidable $200–$400/month in idle charges. Configure JupyterServer idle shutoff to 30 minutes.
• SageMaker Serverless Inference: Remains at $0.00001/GB processed + $0.0000002 per inference, with no minimum charge. Appropriate for low-frequency inference where cold start latency is acceptable (<3 req/second average).

Bedrock Knowledge Bases — New in 2025–2026

Bedrock Knowledge Bases (managed RAG) introduced embedding and retrieval charges in 2025. OpenSearch Serverless (the default vector store) carries a minimum charge of 2 OCUs ($0.24/hour — $172/month minimum) even when idle. Teams building RAG applications with Bedrock Knowledge Bases should factor this floor into their cost model or evaluate alternative vector store options (pgvector on RDS, Pinecone) depending on scale.