Platform

The core Claude API endpoint — you send a list of messages and a model ID, you get a response, and every other platform feature is a parameter on top of it.

Reusing already-processed parts of a prompt across requests, at a tenth of the input price on a cache hit.

Asynchronous processing of large request volumes at a 50% discount on both input and output tokens.

The Start, Build and Scale tiers that govern how many requests and tokens an API organisation can use, and the paths beyond them.

A research-preview mode that trades premium pricing for significantly faster output on Claude Opus 5 and Opus 4.8, on the first-party API only.

Three related controls over how much Claude reasons before answering — and a generational split that means code written for one model may not work on the next.

Claude models resold through AWS, billed by AWS, with AWS-style model IDs and — importantly — AWS's own retirement schedule.

Claude models on Google Cloud's Vertex AI and Agent Platform, billed by Google, with three endpoint types and an independent retirement schedule.

Claude on Azure, billed in Claude Consumption Units through the Azure Marketplace, and — unlike Bedrock and Google Cloud — Anthropic-operated.

Anthropic-operated Claude billed through AWS Marketplace in Claude Consumption Units — AWS procurement with first-party model IDs and lifecycle.

Server-hosted agents with a managed sandbox, billed on both tokens and session runtime rather than tokens alone.

The web interface for the Developer Platform — API keys, usage, spend, rate limits, and a workbench for testing prompts.