.png)
Key takeaways
1
Do not choose a managed AI platform on token price alone. Compare model capability, context requirements, latency, region, deployment type, supporting infrastructure, and integration costs.
2
Token pricing is only one part of the bill. Retrieval infrastructure, logging, networking, provisioned capacity, caching behavior, and agent workflows can materially change total cost.
3
FedRAMP High is a hard compliance filter: evaluate the authorized service, model, endpoint, region, data-residency controls, and applicable compliance configuration for the deployment you are considering.
4
Batch inference, prompt caching, and commitments are not universal discounts. Eligibility and economics vary by model, deployment, region, and service configuration.
Choosing between AWS Bedrock, Vertex AI, and Azure OpenAI requires a platform-level comparison of model fit, compliance scope, integration requirements, and total cost, not token rates alone. For predictable eligible supporting infrastructure, evaluate commitment options separately from managed-model inference charges.
This guide covers what AWS Bedrock, Vertex AI, and Azure OpenAI cost and the surrounding services, deployment choices, caching mechanics, and infrastructure that can make total platform cost materially exceed a simple token-rate estimate.
Google announced FedRAMP High authorization for Generative AI on Vertex AI in March 2025. Current eligibility still depends on the specific Google Cloud configuration and compliance scope.
This guide covers what AWS Bedrock, Vertex AI, and Azure OpenAI cost and the surrounding services, deployment choices, caching mechanics, and infrastructure that can make total platform cost materially exceed a simple token-rate estimate.
Choose your platform in 3 questions
1. Where does your primary cloud workload live today?
- Mostly AWS → AWS Bedrock is your starting point.
- Mostly Azure, Microsoft 365, or Dynamics → Azure OpenAI is your starting point.
- Mostly GCP, BigQuery, or Cloud Storage → Vertex AI is your starting point.
- Genuinely multi-cloud or no strong preference → Continue to question 2.
2. What is your primary workload type?
- Multi-model flexibility, model switching, and AWS IAM integration → AWS Bedrock
- OpenAI model access and Microsoft enterprise integration → Azure OpenAI
- Long-context processing, BigQuery integration, or Gemini workloads → Vertex AI
3. Does compliance override the platform preference?
For FedRAMP High workloads, do not make a platform-wide yes-or-no decision. Validate the exact authorized service, model, endpoint, region, data-residency configuration, and customer authorization boundary.Google announced FedRAMP High authorization for Generative AI on Vertex AI in March 2025. Current eligibility still depends on the specific Google Cloud configuration and compliance scope.
Compliance validation checklist
Before approving a platform for a regulated workload, verify:- Service
- Model
- Endpoint
- Region and data-residency configuration
- Customer authorization boundar
What is actually comparable?
AWS Bedrock vs Vertex AI vs Azure OpenAI pricing snapshot| Platform | Representative model | Deployment and region | Input / 1M | Output / 1M | Caching treatment | Batch rate where supported |
|---|---|---|---|---|---|---|
| AWS Bedrock |
OpenAI gpt-oss-120b, model ID
openai.gpt-oss-120b-1:0
|
Standard, Asia Pacific (Sydney) | $0.1545 | $0.6180 | No separate cached-input rate shown for this Standard rate-card entry | Batch: $0.0773 input / $0.3090 output |
| Google Vertex AI | Gemini 3.5 Flash-Lite | Standard, Global | $0.30 | $2.50 | Cached input: $0.03 / 1M | Flex/Batch: $0.15 input / $1.25 output; cached input $0.015 |
| Azure OpenAI | GPT-5 2025-08-07 | Global Standard | $1.25 | $10.00 | Cached input: $0.13 / 1M | Global Batch: $0.63 input / $5 output; cached input $0.07 |
Sources: Amazon Bedrock pricing, Vertex AI generative AI pricing, and Azure OpenAI pricing.
These numbers show why a raw rate table is only the starting point. For example, the AWS row uses a gpt-oss model in Sydney, the Google row uses a Gemini Flash-Lite model on a Global endpoint, and the Azure row uses GPT-5 on Global Standard.
Their capabilities, model architectures, deployment scopes, context behavior, and quality targets are different, so the prices should not be interpreted as a benchmark ranking.
A useful comparison first groups workloads into premium, balanced, or budget tiers. Within each tier, compare:
This matters because provider pricing pages publish model-specific rates, while your production decision depends on the cost of reaching the required quality and performance level.
Bedrock combines model inference pricing with separately billed services such as Knowledge Bases, Agents, Guardrails, and supporting AWS resources.
Vertex AI combines generative AI usage with separately metered services such as prediction infrastructure, data processing, storage, networking, and other Google Cloud resources.
Azure OpenAI can use pay-as-you-go token pricing or provisioned throughput, while surrounding Azure services such as networking, logging, retrieval infrastructure, and application compute remain separate cost layers.
These numbers show why a raw rate table is only the starting point. For example, the AWS row uses a gpt-oss model in Sydney, the Google row uses a Gemini Flash-Lite model on a Global endpoint, and the Azure row uses GPT-5 on Global Standard.
Their capabilities, model architectures, deployment scopes, context behavior, and quality targets are different, so the prices should not be interpreted as a benchmark ranking.
A useful comparison first groups workloads into premium, balanced, or budget tiers. Within each tier, compare:
- Quality against the same benchmark
- Context length
- Input-to-output token ratio
- Latency target
- Region
- Deployment type
This matters because provider pricing pages publish model-specific rates, while your production decision depends on the cost of reaching the required quality and performance level.
Three different billing structures
AWS Bedrock, Vertex AI, and Azure OpenAI use different billing structures.Bedrock combines model inference pricing with separately billed services such as Knowledge Bases, Agents, Guardrails, and supporting AWS resources.
Vertex AI combines generative AI usage with separately metered services such as prediction infrastructure, data processing, storage, networking, and other Google Cloud resources.
Azure OpenAI can use pay-as-you-go token pricing or provisioned throughput, while surrounding Azure services such as networking, logging, retrieval infrastructure, and application compute remain separate cost layers.
Why this matters:
A valid FinOps comparison has to normalize the cost layers before comparing platforms. The cheapest token rate does not automatically mean the lowest total platform cost.
How to measure caching correctly
Prompt caching is not a single universal discount. Economics vary by provider, supported model, cache-write and cache-read pricing, TTL, storage behavior, and deployment configuration.Before calculating savings, measure:
Uncached input tokens
Cache-write tokens
Cache-read tokens
Output tokens
Cache hits and misses
Cache TTL or storage duration
Retry tokens
Tool-call tokens
A practical cost model is:
Total inference cost = uncached input + cache writes + cache reads + output + cache storage + retries + tool-call usage
Use one disclosed worksheet for all platforms and keep the model version, region, deployment type, and token-accounting assumptions consistent.
Do not estimate caching savings from total token volume alone.
Bedrock Knowledge Bases can use Amazon OpenSearch Serverless, while agentic workflows can create additional model calls for reasoning, tool selection, tool results, and synthesis.
Cross-region inference also needs careful evaluation. AWS states that cross-region inference does not add a separate routing charge. It does have data-residency and pricing-model implications.
Long-context workloads and caching behavior should be modeled using the pricing rules of the specific model being deployed rather than applying one rate across the entire Vertex AI catalog.
Provisioned throughput can be useful for sustained workloads, but teams should validate utilization against production telemetry before committing capacity.
Do not estimate caching savings from total token volume alone.
Where each platform adds cost
AWS Bedrock
AWS Bedrock costs can extend beyond model inference through Knowledge Bases, Agents, Guardrails, supporting storage, testing, retries, and other AWS resources.Bedrock Knowledge Bases can use Amazon OpenSearch Serverless, while agentic workflows can create additional model calls for reasoning, tool selection, tool results, and synthesis.
Cross-region inference also needs careful evaluation. AWS states that cross-region inference does not add a separate routing charge. It does have data-residency and pricing-model implications.
Vertex AI
Vertex AI costs can include model inference, prediction infrastructure, video or image generation, grounding, data movement, storage, and other Google Cloud resources.Long-context workloads and caching behavior should be modeled using the pricing rules of the specific model being deployed rather than applying one rate across the entire Vertex AI catalog.
Azure OpenAI
Azure OpenAI costs can include pay-as-you-go inference or provisioned throughput alongside networking, logging, retrieval infrastructure, application compute, and other Azure services.Provisioned throughput can be useful for sustained workloads, but teams should validate utilization against production telemetry before committing capacity.
Five LLM cost levers
The five optimization areas to evaluate are:
1
Enable prompt caching where the selected model and deployment support it.
2
Right-size model selection for each task instead of defaulting to a flagship model.
3
Evaluate asynchronous batch inference for eligible non-urgent workloads, using the current model- and deployment-specific batch rate.
4
Attribute inference usage to the responsible business unit or application.
5
Benchmark provisioned capacity only after collecting representative production telemetry.
AWS and Google advertise batch discounts for selected supported models, while Azure pricing and support depend on the applicable model and deployment configuration. Always verify the current provider rate before assuming a fixed percentage reduction.
An AWS Compute Savings Plan does not discount Amazon Bedrock model-inference token charges. AWS describes Compute Savings Plans as applying to eligible EC2, Fargate, and Lambda usage.
For eligible supporting infrastructure, AWS states that Compute Savings Plans can provide up to 66% savings compared with eligible On-Demand usage in exchange for a one- or three-year hourly-spend commitment. Actual savings depend on eligibility, usage profile, plan type, region, operating system, tenancy, term, and payment option.
Idle resources still matter
Inference tokens are not the only resources FinOps teams should audit. Review unused endpoints, retrieval infrastructure, provisioned capacity, development resources, fine-tuned deployments, logging, and surrounding compute.Commitments for the infrastructure around AI
Commitment programs must be separated from managed LLM inference pricing.An AWS Compute Savings Plan does not discount Amazon Bedrock model-inference token charges. AWS describes Compute Savings Plans as applying to eligible EC2, Fargate, and Lambda usage.
For eligible supporting infrastructure, AWS states that Compute Savings Plans can provide up to 66% savings compared with eligible On-Demand usage in exchange for a one- or three-year hourly-spend commitment. Actual savings depend on eligibility, usage profile, plan type, region, operating system, tenancy, term, and payment option.
Which AI cost layers may involve commitments?
| Cost layer | Native commitment treatment |
|---|---|
| Managed-model inference | Do not assume eligibility. Check the specific AI service and pricing model. |
| Retrieval or vector infrastructure | Depends on the underlying service and resource. |
| Application compute | May qualify for native compute commitments where the resource is eligible. |
| Networking | Verify service-specific pricing separately. |
| Logging and monitoring | Verify the applicable service pricing separately. |
| Provisioned AI capacity | Product-specific. Do not assume a general compute commitment applies. |
How Usage.ai helps optimize AI infrastructure commitments
We focus on the pricing and commitment layer around eligible cloud infrastructure supporting AI workloads. Once teams establish a stable usage baseline, we analyze commitment opportunities, execute approved commitment-management actions, and continuously optimize coverage as usage changes.Through ourFlex Insured Commitment Program, teams can achieve up to 50% savings on covered cloud spend, on average, across eligible AWS, Azure, and GCP infrastructure while reducing the financial risk of long-term cloud commitments.
Our pricing is based on a percentage of realized savings, and eligible commitments include cashback protection for underutilization.
The cloud providers continue to set the underlying AWS Bedrock, Vertex AI, and Azure OpenAI inference pricing, eligibility rules, and billing mechanics. We optimize the eligible infrastructure and commitment layer around those workloads, helping teams capture commitment savings with less operational overhead and commitment risk.
Best fit and when to reconsider
RAG workloadsBest fit: prioritize data location, retrieval integration, model quality, and total retrieval-plus-inference cost.
Reconsider when: vector infrastructure or data movement becomes a significant part of total cost.
Batch document processing
Best fit: use a supported asynchronous batch option when turnaround requirements allow it.
Reconsider when: latency requirements become interactive or the selected model does not support the required batch deployment.
Interactive copilots
Best fit: prioritize latency, model quality, identity integration, and ecosystem fit.
Reconsider when: traffic patterns or model quality requirements materially change.
Agentic workflows
Best fit: evaluate tool integration together with complete token accounting.
Reconsider when: retries, tool calls, or orchestration materially increase usage.
Long-context analysis
Best fit: compare models using the actual context length and caching behavior required by the workload.
Reconsider when: context size, caching strategy, or output requirements change.
Frequently asked questions
What is the difference between AWS Bedrock, Vertex AI, and Azure OpenAI pricing?
All three offer managed AI inference, but their model rates, deployment options, caching rules, batch pricing, and surrounding cloud costs differ. Compare the same workload, capability tier, region, deployment type, context requirements, and token-accounting assumptions before drawing a cost conclusion.
Is Vertex AI cheaper than AWS Bedrock?
Not universally. The answer depends on the selected models, quality requirement, token mix, context length, deployment, region, and supporting infrastructure. Compare equivalent workload tiers rather than isolated published token rates.
When does provisioned AI capacity make sense?
Provisioned capacity should be evaluated against representative production telemetry. It is most relevant when sustained workload demand can use the committed capacity efficiently.
How does prompt caching reduce LLM costs?
Caching can lower the cost of repeated input where the selected provider, model, and deployment support it. Calculate savings from actual cache writes, reads, hits, TTL behavior, uncached input, and output rather than applying a universal discount.
Can cloud commitments reduce AI costs?
They can reduce eligible supporting infrastructure costs, but commitment eligibility depends on the specific resource. Do not assume a general compute commitment discounts managed LLM inference tokens.
Share
Facebook
X
LinkedIn
Reddit
