Quick answer
Production AI on AWS costs $800/month for a focused chatbot to $45,000/month for an enterprise multi-agent system, based on 24 builds we shipped 2024-2026. The published list-price quotes nearly always miss three line items that bust budgets: NAT Gateway data processing fees ($600-$2,400/mo), inter-AZ data transfer ($0.01-$0.02/GB compounding), and CloudWatch Logs ingest above the free tier. We have audited 18 invoices where the actual monthly bill ran 2-4x the quoted number.
The four cost tiers from 24 production builds
| Tier | Use case | Monthly AWS | Median inference |
|---|---|---|---|
| Focused chatbot | Single workflow, <500 req/day | $800 - $2,400 | Bedrock Claude $0.02-$0.08 per req |
| Multi-flow agent | 3-5 tools, 1K-5K req/day | $2,400 - $9,000 | $0.05-$0.18 per req |
| Full ops agent | 5-10 tools, 5K-25K req/day | $9,000 - $22,000 | $0.08-$0.24 per req |
| Enterprise multi-agent | 10+ agents, 25K+ req/day | $22,000 - $45,000 | $0.12-$0.40 per req |
The 3 line items that turn $5K quotes into $23K invoices
- NAT Gateway data processing — $600-$2,400/month. Every byte going through a NAT Gateway is charged at $0.045/GB. AI agents calling external APIs (Anthropic, OpenAI, third-party tools) through a private VPC hit this on every request. Most quotes show the $0.045/hour NAT Gateway fee but forget the processing fee.
- Inter-AZ data transfer — $0.01-$0.02 per GB, both directions. Multi-AZ Bedrock + multi-AZ RDS + multi-AZ EKS = three round trips per request, each charged. On a 25K-request-per-day workload this can add $1,200-$3,600/month that does not show up in any "what does Bedrock cost" calculator.
- CloudWatch Logs above free tier — $0.50 per GB ingested. AI agents log heavily for eval pipelines. A multi-agent system can ingest 30-80 GB/month of structured logs. That is $15-$40/month per agent. Sounds small until you have 10 agents.
Getting hit by surprise AWS bills? Free 30-min AWS audit. We look at your last 3 invoices + Cost Explorer view, identify the top 3 cost leaks, and give you a written remediation plan.
Book a free AWS audit →Bedrock vs SageMaker vs EC2 GPU — when to use which
- Bedrock — pay-per-token, no infrastructure to manage. Best for <25K requests/day. Median cost ~$0.06 per request on Claude Sonnet 4.7. Wins on operational simplicity for <90% of use cases.
- SageMaker real-time endpoints — host your own fine-tuned model. Best when you have a custom model that delivers better-per-dollar accuracy than Bedrock. Median fixed cost $850-$1,800/month for a ml.g5.xlarge endpoint. Break-even vs Bedrock typically at 50K+ requests/day.
- EC2 GPU instances — full self-host. Best when you have ITAR/GovCloud constraints or specific GPU needs Bedrock cannot meet. Highest operational overhead. We recommend this only when forced by constraints, not by cost preference.
Real bill: $5,000 quote vs $22,800 actual (client engagement March 2025)
A US e-commerce client came to us after their first month on AWS Bedrock hit $22,800 against a vendor quote of $5,000. Breakdown of the gap:
| Line item | Quoted | Actual |
|---|---|---|
| Bedrock inference (Claude) | $4,200 | $11,400 |
| NAT Gateway data processing | $0 (missing) | $2,800 |
| Inter-AZ data transfer | $0 (missing) | $3,400 |
| CloudWatch Logs | $0 (missing) | $1,900 |
| RDS Postgres (eval store) | $400 | $1,200 |
| S3 + CloudFront | $200 | $680 |
| EC2 (eval runner) | $200 | $420 |
| Total | $5,000 | $22,800 |
After our audit + remediation (Bedrock prompt caching, NAT Gateway → VPC endpoints for AWS-internal traffic, CloudWatch Logs retention reduced, RDS reserved instance) the monthly bill dropped to $8,200. Same workload. Different architecture.
FAQ
Should we use Bedrock or call Anthropic directly?
Bedrock for any US deployment with compliance constraints (HIPAA BAA, SOC 2 audit-log capture). Direct Anthropic API for everything else — usually 8-15% cheaper at scale due to no AWS markup. We benchmark both per engagement.
How do we cut AI costs on AWS?
9 levers, ranked: prompt caching (Bedrock + Claude direct), Spot for non-prod, VPC endpoints to eliminate NAT data fees, S3 Intelligent-Tiering, RDS Reserved Instances, CloudWatch retention windowing, model right-sizing (Haiku vs Sonnet vs Opus), eval-set sampling, and old-snapshot deletion. Average savings across 18 audits: 40-65%.
When does SageMaker beat Bedrock on cost?
Above 50K requests/day for a workload where you have fine-tuned a smaller model that performs comparably. Below that, Bedrock pay-per-token nearly always wins.
Do you offer ongoing AWS cost management?
Yes — monthly cost review + remediation as part of our SLA tier. Typical clients see $4,200/month saved on average within 6 weeks of the first audit.
Free audit, 48-hour delivery
Stop bleeding cash on AWS
Send us your last 3 AWS invoices + your Cost Explorer view. We come back in 48 hours with the top 3 cost leaks, the projected savings, and a fixed-price remediation quote if you want us to ship the fix. Average savings across 18 audits: $4,200/month.
Related resources
- 9 cost levers that cut AI bills 40-65%
- AI on AWS Bedrock — full architecture guide
- Case study: HIPAA AI rebuild on AWS Bedrock
Methodology
Cost ranges and quote-vs-actual data from 24 Braincuber AI on AWS engagements shipped 2024-2026 and 18 documented client AWS cost audits. AWS Bedrock + Claude Sonnet 4.7 pricing referenced as of April 2026 public list rates. Inter-AZ data transfer pricing per AWS Data Transfer pricing page. Client invoice details anonymized; specific dollar amounts published with permission. The $4,200/month average savings figure is the median across 18 completed audits, not a maximum.
Founder and CEO of Braincuber. Has scoped and shipped 500+ Odoo, AI, and cloud projects for US mid-market and global brands. Takes every founder call personally — no SDR layer between buyers and the people building the system.