If you estimate Azure OpenAI or Amazon Bedrock costs from the headline price, your forecast is probably about 10% low. Keeping requests in one geography costs more than letting them run anywhere, and most teams with data-residency rules pay that uplift on every call.
Azure OpenAI: Global, Data Zone and Regional
Azure OpenAI deployments come in types. Global deployments can process requests in any Azure region and carry the list price. Data Zone deployments keep processing within the United States or the European Union, and Regional deployments within one region. Both cost more.
| Per million tokens | Global input | US or EU Data Zone input | Global output | Data Zone output |
|---|---|---|---|---|
| gpt-5 | $1.25 | $1.375 | $10.00 | $11.00 |
| gpt-4.1 | $2.00 | $2.20 | $8.00 | $8.80 |
| gpt-4.1-mini | $0.40 | $0.44 | $1.60 | $1.76 |
| gpt-4o-mini | $0.15 | $0.165 | $0.60 | $0.66 |
Across these models the Data Zone price is exactly 10% above Global, for input, output and cached input alike.
Amazon Bedrock: cross-region inference profiles
On Bedrock, Claude models are called through inference profiles. The global. profile can route anywhere and has the base price. Geographic profiles such as us., eu., jp. and au. keep requests in that geography and cost more. AWS GovCloud costs more again.
| Claude Sonnet 4.5 on Bedrock, per million tokens | Input | Output | Uplift |
|---|---|---|---|
| global. profile | $3.00 | $15.00 | — |
| us., eu., jp. or au. profile | $3.30 | $16.50 | +10% |
| us-gov profile (GovCloud) | $3.60 | $18.00 | +20% |
What it does to a budget
A team that forecasts $20,000 a month from list prices, but runs everything on EU Data Zone deployments for GDPR reasons, will see about $22,000. Over a year that's $24,000 nobody planned for, and the provider invoice won't say why: it shows totals per resource, not the price each call was charged at.
What to do about it
- Use Global where you can. If a workload has no residency requirement, a Global deployment or
global.profile is the cheapest option. - Keep residency where you need it, and budget for it. Add 10% for Data Zone and geographic profiles and 20% for GovCloud.
- Price each call for where it ran. Record the deployment or profile with every call, so cost per service and per customer uses the right rate.
Alye does the last step for you: it reads the endpoint and model id of each call, prices it for that Azure deployment type or Bedrock profile, and labels where it ran. See Azure OpenAI cost tracking and Bedrock prices by profile.
Prices are list prices on the date of this post, from our model price list. Check Azure's and AWS's pricing pages for current rates.