Azure OpenAI vs Bedrock vs Vertex AI: what changes the bill

By Rohit Athaide · Published 4 Nov 2026

Short answer: several large models are sold in more than one place: directly from the model maker, and through the big clouds. The model is the same, but the bill is not always. Price per token, discount rules, capacity options, regions and extra services can all differ, so compare the exact model and version on each platform.

Key points

• Compare the same model name and version, never different models.
• Check price per million input and output tokens on each platform.
• Reserved capacity bills by the hour, used or not.
• Batch and caching discounts vary by platform.
• Extras around the model add to the bill.

Which platforms are there?

Each hosts a different set of models, and the set changes.

What can differ for the same model?

1. Price per million tokens. Often equal to the maker’s list price, but not always. Check the platform’s page.
2. Pricing modes. Pay per token, or reserve capacity for steady throughput. Reserved capacity bills for the hours, used or not.
3. Batch and caching discounts. Rates for batch jobs and cached prompts vary by platform. See the discounts page.
4. Regional and data residency options. Some regions or in-country processing cost more.
5. Extra services. Content filters, guardrails, knowledge bases and logging are billed on top.
6. Quotas. Default rate limits differ and may need a request to raise.

What does a worked example look like?

Made-up round numbers to show the method, not real prices. A workload sends 100 million input tokens and receives 20 million output tokens a month.

The token prices match and the bills do not, because Platform Q bills a safety feature per call that Platform P includes or leaves off. The guardrail line here is exaggerated to make the point: the model price is rarely the only line. Now try reserved capacity. Suppose a reserved unit costs 2,000 a month and carries up to 150 million tokens, and per-token pricing averages 12.50 per million. At 120 million tokens, per-token costs 1,500 and the reserved unit loses. At 150 million it costs 1,875 and still loses, and at 180 million it would cost 2,250, but that load needs a second unit. Reserved capacity pays off only when the workload fills it close to full all month.

What sits around the model?

A model call is rarely the only cost. Search indexes, vector stores, storage, and data transfer between services add to it. See RAG chatbot monthly cost and vector database and embedding costs.

Which mistakes are common?

• Comparing a model on one platform with a different model on another.
• Ignoring output tokens, which usually cost several times more than input tokens.
• Reserving capacity before measuring traffic.
• Forgetting that safety and logging features are billed.

How do I compare?

1. Pick the exact model and version.
2. Estimate input and output tokens per request, and requests per month.
3. Price them on each platform, with caching and batch rates if you will use them.
4. Add the platform extras you need.

Decision checklist

• Same model and version on each platform?
• Are input and output tokens priced separately?
• Is reserved capacity justified by measured volume?
• Are extras and logging counted?

Next steps

Our AI cost estimator does the first three steps from published prices, and the AI methodology page says where the numbers come from. For the basics of tokens, read AI model token pricing explained, and if volume is very high, see self-hosting an LLM versus paying per token.