Anthropic API — AI Glossary

Reference details

Related terms

• OpenAI API
• Claude
• Anthropic Python SDK

How it works

1. The system instruction is a top-level parameter rather than an element of the message list, so the conversation array contains only user and assistant turns.
2. Those turns must alternate and begin with a user turn, which means a client assembling history has to merge consecutive same-role messages rather than appending freely.
3. Message content is a list of typed blocks rather than a string — text, image, tool use, tool result — which is what lets a single turn carry an image and a question together as one coherent input.
4. Tool results are themselves content blocks inside a user turn referencing the originating call's identifier, so the loop is expressed in the message structure rather than in a separate role.
5. Because the assistant turn is just another message, a partially written one can be supplied and the model continues it, which is how output format is constrained by prefilling rather than by instruction.
6. A stop reason distinguishes a completed turn from a token limit from a tool call, and a response that ends on a tool call is the signal that the caller must execute and return a result before the turn is finished.

Example: The system prompt is a field, not a message

The most common porting bug is structural. In the chat completions convention the system instruction is the first element of the messages array with role “system”. Here it is a separate top-level `system` parameter, and messages alternate user and assistant. Code that moves across without changing that either drops the instruction or sends it as a user turn — and the failure is silent, because the request is still valid and the model still answers, just without the constraints you thought you had set. Compatibility layers usually handle the translation; hand-rolled ports frequently do not.

Commonly confused with

It is not natively an OpenAI-compatible endpoint — the request shape, the system-prompt placement and the tool-use format all differ, and compatibility comes from a translation layer such as a gateway. Vendor-hosted deployments add another axis: the same model family reached through a cloud marketplace may differ in available features and region.

When to use it

Reach for it when:

• Direct integrations where you want the native feature surface without translation loss
• Long-context, vision and extended-thinking work, which are first-class here
• Tool use where the native format avoids a lossy conversion

Reach for something else when:

• Dropping into code that assumes the chat completions shape without adapting it
• Assuming parameter names and defaults carry over from another provider
• Pinning to a floating model alias in production instead of a dated version

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Anthropic API — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path