AWS Bedrock regional pricing varies by model, with multipliers from 0.96x to 1.70x

The core finding

Cloud cost models typically treat an AWS region as a multiplier applied to a us-east-1 base rate. An analysis posted on dev.to by AI Change Watch tested that assumption against Amazon Bedrock's full on-demand catalogue and found it does not hold.

The author read all 1,586 of Bedrock's on-demand, in-region token meters and compared each with its us-east-1 equivalent. Of those, 1,424 meters across 32 regions were directly comparable, and together they produce 123 distinct multipliers. If region functioned as a simple uplift, there would be at most 32.

The distribution is lopsided. Nearly a third of the meters (30.3%) cost exactly what they do in Virginia, which puts the first quartile at ×1.000. The median is ×1.183 and the third quartile ×1.200, but the extremes run from ×0.964 to ×1.700. The top of the range is São Paulo, where Mistral 7B output tokens cost $0.34 instead of $0.20. Four meters are actually cheaper outside us-east-1 — the input and output prices of two Nova Sonic models in us-west-2, at ×0.967 and ×0.964 — so the uplift does not even have a floor at 1.

One region, many markups

London (eu-west-2) is the clearest illustration. Across both input and output directions the region has 94 comparable meters and 25 distinct multipliers. Two models — Kimi K3 and Grok 4.7 — bill input at exactly the Virginia price, at $3.300 and $2.200 respectively. Meanwhile Llama 3 8B input carries a ×1.300 markup ($0.300 to $0.390), Mistral 7B ×1.333, Nova Pro ×1.412 ($0.800 to $1.130), and Ministral 3B 3.0 ×1.600 ($0.100 to $0.160). Those examples cover nine of London's 47 comparable input meters.

The upshot is that whether a blanket claim like "London costs 55% more" is accurate depends entirely on which model you happened to pick. Consistency elsewhere is rare: only 9 of the 32 regions apply a single multiplier, and seven of those nine have just four meters each, which the author calls a sample rather than a policy. The two consistent regions with meaningful coverage are us-east-2 (129 meters, all ×1.000) and me-central-1 (13 meters, all ×1.000). In total, 23 of 32 regions price different models differently relative to Virginia.

The Azure contrast

Azure organises pricing along a different axis, according to the same analysis. A Global Standard deployment has one price, served for every region Azure lists. Cost only enters when a model is pinned to a specific geography: regional pinning has a median multiplier of ×1.100 with a range of ×1.100 to ×1.320 across 10 distinct multipliers, while data zone pinning has a median of ×1.100 with a range of ×1.100 to ×1.200 across 6 multipliers.

Azure is not perfectly uniform either — 18 of the 21 models with regional pricing do vary between regions. But 10 distinct multipliers against AWS's 123, over a comparable number of meters, is something the author characterises as a structural difference rather than a matter of degree.

A methodology warning

The comparison covers only Bedrock's on-demand standard tier with in-region placement; Batch, Flex and Priority tiers, plus cross-region inference, sit in separate tables and were excluded. The author says that exclusion was learned the hard way: a first pass that did not pin the deployment type produced a tidy headline — that Nova 2.0 Lite cost 6.16× more in some regions — which turned out to compare Flex pricing in one region against Priority in another. The underlying rows were real; the comparison answered a question nobody asked. Prices were read from the vendor's published tables, dated 11 October 2026, and never derived from a reference figure.

Why it matters

If you budget Bedrock spend by applying a regional uplift to a us-east-1 price, this analysis suggests the result can be wrong by a wide margin in either direction, because no single factor is right for more than a slice of the catalogue. The practical guidance is to look up the price for the specific model, direction and region you actually call, and to check your two heaviest meters rather than trusting the ×1.183 median — a bill is a weighted sum over the particular meters you hit, and those range from ×0.964 to ×1.700. On Azure, the first question is the deployment type: under Global Standard, region is not a cost axis at all, and a region migration changes nothing on the invoice, which is worth knowing before committing engineering time to one.