Model Selection Guide

Agents consistently get these wrong. Four patterns:

Always look up current model IDs: aws bedrock list-foundation-models --region <region> and aws bedrock list-inference-profiles --region <region>, or refer to the latest Bedrock supported models (opens in a new tab).

Critical: Some models do not support on-demand invocation with base model IDs and require an inference profile ID instead. Before using a model, check aws bedrock list-inference-profiles --region <region> — if an inference profile exists for the model, use the inference profile ID. If you get ValidationException: on-demand throughput isn't supported, switch to the inference profile ID.

Model Access Provisioning

Most serverless models are automatically available without manual enablement. Use IAM policies and SCPs to control which models can be used.

What still requires action:

• Anthropic models: Enabled by default but require a one-time usage form submission before first use (via Bedrock console playground or PutUseCaseForModelAccess API). For AWS Organizations, submitting via API at the management account level extends approval to child accounts.
• Third-party Marketplace models: A subset of models require AWS Marketplace subscription, which is created automatically on first invocation if the caller has aws-marketplace:Subscribe permission.
• EULAs: Some models still require EULA acceptance. Review EULAs at the model card in Model Catalog (opens in a new tab) or the Bedrock third-party model terms (opens in a new tab).

Access control: Use IAM policies (bedrock:InvokeModel scoped to specific resource ARNs) and SCPs to control which models can be used. Use bedrock:ListFoundationModels for listing models and bedrock:GetFoundationModel for getting details about a specific model. The IAM Resource ARN format depends on the model ID type:

• Inference profile ID → arn:aws:bedrock:<region>:<account-id>:inference-profile/<profile-id>
• Base model ID → `arn:aws:bedrock:::foundation-model/`````
• These are different ARN formats and are not interchangeable. See Bedrock IAM resource types (opens in a new tab)
• For least-privilege policies scoped to specific inference profiles, you MUST include BOTH the inference profile ARN (arn:aws:bedrock:<region>:<account-id>:inference-profile/<profile-id>) AND the foundation model ARN with a wildcard region (arn:aws:bedrock:*::foundation-model/<model-id>), because the request may be routed to any region in the profile – otherwise bedrock:InvokeModel calls fail with AccessDeniedException. See Prerequisites for inference profiles (opens in a new tab)

INVALID_PAYMENT_INSTRUMENT error: Some AWS accounts (especially Organizations with European billing/SEPA) get this error when subscribing to Marketplace models. This is an account billing issue, not a Bedrock issue.

• Workaround: temporarily set a VISA/credit card as default payment method
• Alternative: per AWS re:Post user reports, adding USD payment profiles in the organization management account (Billing → Payment Preferences → Payment profiles) for service providers ending with “- Marketplace” may resolve the issue
• Contact AWS Support if the issue persists

Selection Criteria

List models with capabilities: aws bedrock list-foundation-models --region <region>

Quick defaults (verify current availability — new models are added frequently, check aws bedrock list-foundation-models --region <region> or the Bedrock supported models page (opens in a new tab)):

• General purpose / reasoning: Claude Sonnet
• Fast + cheap: Claude Haiku or Nova Micro
• Open-source / fine-tuning: Llama
• Multilingual: Cohere Command or Claude
• Code generation: Claude Sonnet or Llama

Decision framework — choose based on:

Embedding Models for Knowledge Bases

This is a non-obvious choice that affects KB quality. The table below shows common options — additional embedding models (including multimodal embeddings) are available. Check aws bedrock list-foundation-models --by-output-modality EMBEDDING --region <region> for the current list.

Critical: The embedding model dimensions MUST match the vector store index dimensions. Mismatched dimensions cause ingestion failure.

Refer to the latest AWS documentation on Bedrock embedding models for current options.

Pricing Models

Refer to the latest AWS documentation on Bedrock pricing for current rates and discount percentages. Pricing changes without notice — do not hardcode pricing assumptions.