The Claude 3P Glossary

Bookmark this page and share it with whoever joins your Claude evaluation next — procurement, legal, security, or a skeptical engineering manager. Each entry is two or three sentences, not a lecture. Terms are grouped so you can read the section relevant to today's meeting and skip the rest.

Core concepts

Inference

Running a trained model to get an answer. When your application sends a request to Claude and gets a response back, that is inference. You pay for inference; you never train the model yourself.

Token

The unit text is broken into for processing and billing — a short chunk of characters, often part of a word. Both your input and Claude's output are measured and priced in tokens.

Context window

The maximum amount of text (in tokens) a model can consider in one request. Claude Opus 4.8 and Sonnet 5 have 1M-token windows; Claude Haiku 4.5 has 200K.

Prompt

Everything you send to the model in a request: instructions, documents, examples, and the user's actual question.

System prompt

Standing instructions your application always sends — tone, rules, role — separate from what the end user types. It is where most of your app's "personality" and guardrails live.

Model ID

The exact string your code uses to name a model, such as claude-sonnet-5. On Amazon Bedrock the same models carry an anthropic. prefix (for example anthropic.claude-sonnet-5); everywhere else the bare ID is used.

Hallucination

A confident-sounding but wrong answer. A known failure mode of all large language models, managed with grounding (see RAG), review steps, and testing — not eliminated by any vendor.

Platforms and where Claude runs

1P (first-party)

Claude bought directly from Anthropic — the Claude API at anthropic.com.

3P (third-party platforms)

Claude accessed through a cloud provider instead of directly from Anthropic. There are four surfaces: Claude Platform on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.

Amazon Bedrock

AWS's managed service for foundation models, operated by AWS. Claude is one of the models it offers. Access and billing run through your AWS account and IAM.

Google Vertex AI

Google Cloud's machine-learning platform. Claude runs inside your GCP project, authenticated with Google's standard Application Default Credentials.

Microsoft Foundry

Microsoft's Azure-based platform for AI models. You create a Foundry resource, get an API key, and call Claude through it. Feature support is broad, though some features are still in beta as of July 2026.

Claude Platform on AWS

The newest option: operated by Anthropic but running on AWS, inside your AWS boundary. It gets new API features the same day they ship on the first-party API.

Region

The geographic location of the cloud infrastructure serving your requests. Region choice affects latency and data-location questions your legal team will ask.

Global endpoint

A routing option (notably on Vertex AI, where you can set region="global") that lets the platform serve your request from available capacity rather than pinning it to one region.

API features

Messages API

The core request/response interface for Claude: you send a list of messages, you get a reply. Available on all four platforms.

Streaming

Receiving the response word by word as it is generated, instead of waiting for the whole answer. Standard for chat-style interfaces; available everywhere.

Tool use

Letting Claude call functions you define — look up an order, query a database — so answers are grounded in your systems. Available on all four platforms.

Vision

Sending images (scans, screenshots, photos of forms) as part of a request, not just text. Available on all four platforms.

Adaptive thinking

Also called extended thinking: the model spends more reasoning effort on hard problems before answering. Available on all four platforms.

Prompt caching

Reusing a long, repeated portion of your prompt (a policy manual, a schema) across requests so you do not pay full price to resend it every time. Available on all four platforms and usually the single biggest cost lever.

Batch API

Submitting large volumes of non-urgent requests for asynchronous processing. Not available on Bedrock, Vertex AI, or Foundry; available on Claude Platform on AWS.

Integration and operations

RAG (retrieval-augmented generation)

Fetching relevant documents from your own content first, then including them in the prompt so Claude answers from your material rather than from memory.

Rate limit / quota

The ceiling on how many requests or tokens you can push through per unit of time. Each platform meters differently; raise limits before launch day, not after.

SDK

The official client library your developers install (the anthropic package in Python) that handles each platform's authentication and request format, so most application code stays identical across platforms.

Where to go next

With the vocabulary settled, read What Is Claude 3P? for the five-minute overview, or go deeper on billing units in Tokens 101. The platforms section of the main guide compares all four surfaces side by side.