BedrockExplorer

Methodology

Snapshot Sep 15, 2026 · source aws · schema v1

Sources

  • Model catalogue & availability — bedrock:ListFoundationModels in each of the 31 regions tracked here, plus ListInferenceProfiles. A model is shown as On-demand on-demand when it reports ON_DEMAND, and Inference profile only when it only reports INFERENCE_PROFILE, or when a profile hosted in that region routes to it.
  • Pricing — the AWS Price List Query API (AmazonBedrock), normalised to USD per 1M tokens. The API mixes per-1K and per-token units; both are converted.
  • Quotas — Service Quotas, matched to models by name. Quota names are prose with no model ID, so anything that does not match confidently is dropped rather than attached to the wrong model.

What AWS does not publish

Three things this site shows cannot be read from any AWS API, and are hand-maintained:

  • Context window and max output tokens. No Bedrock API returns them.
  • Tool-use support. Not exposed either.
  • Anthropic pricing. Of the ~11,000 Bedrock usage types in the Price List API, exactly ten are Anthropic — all legacy (Claude 2.0/2.1, Claude 3 Sonnet/Haiku, Instant), input-tokens only, us-east-1 and us-west-2 only. Every modern Claude model has zero SKUs. Amazon Nova, Meta, Mistral and most others are fully covered, so they are left to the API.

Hand-maintained prices are tagged curated in every table and carry a verification date. Cache and batch rates for those models are derived from the base rate using multipliers AWS publishes for models it does price (cache read 0.1×, cache write 1.25×, batch 0.5×), confirmed against Bedrock's own Claude 3.5 Sonnet v2 row.

How the residency verdicts are derived

Data residency asks a different question from the matrix — not whether a model is available in a region, but where the inference runs when you call it. Three inputs, two of them from the API:

  • On-demand availability per region, from ListFoundationModels. A bare model ID addressed at a region is served by that region, so this is the only case reported as staying in region.
  • Declared profile destinations, from the model ARNs that ListInferenceProfiles returns. A model reachable only through a profile is reported as staying inside a jurisdiction only when every destination that profile declares sits inside it.
  • Region → country, hand-maintained. AWS's own geography field is too coarse: "Europe" contains the EU, the United Kingdom and Switzerland, and eu-west-2 (London) and eu-central-2 (Zurich) carry an eu- code without being EU member states.

Global profiles are reported as unbounded, not as narrow. The API names a single region for them while AWS documents global routing as worldwide; both cannot be a complete destination list, so the site declines to certify them in either direction rather than trusting the narrower answer. The same reflex applies to a region missing from the country table: it counts toward no jurisdiction instead of defaulting into one.

This describes where inference executes according to what AWS declares. It is not legal advice, and it says nothing about logs, retention, model-improvement terms or where your own application tier runs.

Deliberate omissions

  • Provisioned-throughput-only models. Bedrock lists context variants (…-v1:0:24k) as separate model IDs that support only PROVISIONED. They are excluded: they are a different purchasing model, and showing an on-demand price against them would be a number you cannot pay.
  • Bedrock Marketplace. The Marketplace catalogue — DeepSeek distills, Mistral NeMo, Liquid, Arcee and roughly 150 others — is reachable from the Bedrock console but is not returned by ListFoundationModels. Those models run on a SageMaker endpoint you deploy and bill per instance-hour (ml.p5e.48xlarge and the like), not per 1M tokens. Price, region flag and quota would all mean something different for them, so they are a separate product rather than a gap in this table.
  • Regions the build could not read. The pipeline probes every commercial region where AWS declares Bedrock, and a region reaches the table only once at least one model is callable there. A region that errored is dropped the same way — so it is named here rather than left to look like a region Bedrock does not serve.
    • me-central-1 Middle East (UAE) — the region did not answer in time
    • me-south-1 Middle East (Bahrain) — the region did not answer in time
  • Flex and priority tiers. Only the standard latency tier is shown. Flex and priority are separate products at different rates.
  • Non-token pricing. Image, video and embedding models that bill per image or per second show no token price rather than a misleading one.
  • The regional (CRIS) premium. Prices shown are the base published rate. Regional inference profiles (us., eu.) cost about 10% more than global. for Sonnet 4.5 and later.

Refresh cadence

A GitHub Action runs the pipeline daily, commits the snapshot only if it changed, and triggers a redeploy. The site build itself never calls AWS — it reads only data/bedrock-snapshot.json, which is downloadable.

Warnings from the last run

  • 43 models have no price in any region (e.g. nvidia.nemotron-nano-12b-v2, anthropic.claude-sonnet-4-20250514-v1:0, stability.stable-creative-upscale-v1:0)

Accuracy

This is a reference, not a quote. Prices change, regions launch, and the curated overlay can drift between reviews. Confirm in the AWS console before committing spend.