Overview

LLM inference services are available through several commercial cloud platforms and can be accessed through CloudBank:

• AWS provides AWS Bedrock and AWS Bedrock Mantle
• Azure offers Azure AI Foundry
• Google Cloud Platform features Google Agent Platform and Google AI Studio

Because these services differ in which models they support and how you access them, this documentation walks through each service's specific capabilities and access methods.

If you're unsure which model to use, artificialanalysis.ai/models/recommend can help you decide.

Supported Models

¹ CloudBank has access to OpenAI and Anthropic models on AWS; however, due to restrictions by AWS, CloudBank AWS accounts may not be able to access these models. If you need immediate access to a frontier LLM, we recommend using Azure or Google Cloud Platform instead of AWS. Model availability for a given AWS account may depend on regional factors, payment history, and usage patterns, and is subject to change over time.

Programmatic Access Methods

CloudBank's LLM services can be reached five different ways, summarized below. Each method is demonstrated with runnable Python examples in the cloudbank-llm-examples repository.

OpenAI API

Uses an OpenAI-compatible API key to call the OpenAI Chat Completions endpoint via the OpenAI SDK for Python (openai/openai-python), set through OPENAI_API_KEY / OPENAI_BASE_URL. AWS Bedrock Mantle and Azure AI Foundry each expose an OpenAI-compatible endpoint, so the same client and key format work across both — just point OPENAI_BASE_URL at the provider's endpoint.

Bedrock API Keys

AWS-specific. Uses AWS Bedrock's short- or long-term Bedrock API keys (AWS_BEARER_TOKEN_BEDROCK) to call Bedrock's native APIs directly: the Anthropic SDK for Python's AnthropicBedrock client (anthropics/anthropic-sdk-python) for Anthropic models, or the AWS SDK for Python (boto/boto3) Bedrock client.

Google Gen AI API Keys

Uses a Google AI Studio API key (GEMINI_API_KEY) with the Google Gen AI SDK for Python (googleapis/python-genai) to call the Gemini API directly.

OAuth / User SSO

Authenticate once with the cloud provider's own CLI (AWS: aws login; Azure: az login; Google: gcloud auth application-default login), then let each SDK pick up those credentials automatically — no key to store or rotate. Depending on the provider and model family, this works with the OpenAI SDK, the Anthropic SDK (AnthropicBedrock, AnthropicFoundry, AnthropicVertex), the Google Gen AI SDK, or the AWS SDK.

Key / Secret Pairs

AWS only. Uses traditional long-lived AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY credentials, picked up automatically by the AWS SDK for Python (boto3) and by the Anthropic SDK's AnthropicBedrock client, to call Bedrock's Anthropic and open-weights models.

API key vs. OAuth / SSO

API keys are the simplest option and work well for automation — scripts, CI pipelines, or any context with no user present to log in interactively. The tradeoff is that a key is a long-lived secret: it has to be stored securely, rotated, and revoked individually if it leaks.

OAuth / User SSO avoids managing a separate secret at all — it reuses your existing login (aws login, az login, gcloud auth application-default login) and issues short-lived credentials that refresh automatically. Access is tied to your own identity, so it's easier to audit and to revoke (just disable the user) than tracking down a stray key. The tradeoff is that it needs an interactive login step, so for unattended automation you'd use a service account or managed identity instead.

As a rule of thumb: use OAuth/SSO for local development, and API keys (or their automation-friendly equivalents, like AWS Key/Secret Pairs or a service account) for anything that runs unattended.

¹ Short-term keys only work in the region they were created in.