Claude Platform API

API surface by Anthropic · site · lifecycle docs

The Claude Platform is Anthropic’s API for accessing the Claude models. It is designed for developers building applications that use their models, and can be considered “lower-level” than that of Anthropic’s client apps such as Claude.ai and Claude Code.

This page assumes the reader is familiar with the API, as it aims to supplement the official documentation.

Features

This section covers features that aren’t mentioned in the features overview page or the Models API.

• Sampling params: temperature, top_p, top_k¹
• Prefill: Allows “putting words in Claude’s mouth” - prepending text in the assistant message that the model will continue.²

Model feature support matrix

Generated from each model page’s stored capability facts (Models API + cited observations).

Non-public features

• logprobs: Provides log probabilities of output tokens. UK AISI mentioned replying on this non-public API feature for environment realism methods in their report on verbalized eval-awareness on Mythos 5.³
• Steering API: Research preview in 2024 that allowed steering a subset of Claude’s internal features.

System prompt and ANTML

Angle brackets are replaced with curly brackets in this section.

Tools

Anthropic partially documents how tool definitions are provided to models,⁴ and users have extracted other components of it.⁵⁶

In this environment you have access to a set of tools you can use to answer the user's question.
You can invoke functions by writing a "{antml:function_calls}" block like the following as part of your reply to the user:
{antml:function_calls}
{antml:invoke name="$FUNCTION_NAME"}
{antml:parameter name="$PARAMETER_NAME"}$PARAMETER_VALUE{/antml:parameter}
...
{/antml:invoke}
{antml:invoke name="$FUNCTION_NAME2"}
...
{/antml:invoke}
{/antml:function_calls}

String and scalar parameters should be specified as is, while lists and objects should use JSON format. Note that spaces for string values are not stripped. The output is not expected to be valid XML and is parsed with regular expressions.

Here are the functions available in JSONSchema format:
{{ TOOL DEFINITIONS IN JSON SCHEMA }}
{{ USER SYSTEM PROMPT }}
{{ TOOL CONFIGURATION }}

In some leaks, these are followed by {function_results} and {antml:thinking}...thinking about results{/antml:thinking}. This may vary for different model versions.

On claude.ai, the {{ USER SYSTEM PROMPT }} slot is where the app’s own prompt lands — the layer whose published portion is tracked on the system prompts page. The CL4R1T4S captures cited above are pulled from claude.ai, so they show both layers at once: this ANTML scaffolding plus the full app prompt, including the sections Anthropic doesn’t publish (tools, memory_system, search instructions — see unofficial captures).

Reasoning effort flag

The API’s effort request parameter.

{antml:reasoning_effort}40{/antml:reasoning_effort}

Values: low (10), medium (20), high (40), xhigh (120), max (85).

Interleaved thinking

Interleaved thinking the model to think between tool calls and reason after receiving tool results. This is a beta header for extended thinking, automatically enabled with adaptive thinking.⁷⁸

{antml:thinking_mode}interleaved{/antml:thinking_mode}

Value is either auto or interleaved.

Context awareness

Specific models have tokens injected for tracking context window budgets.⁹

The system prompt of every request gives the model its total context window, for example:

<budget:token_budget>200000</budget:token_budget>

After each tool call, the model is given an update on its remaining capacity.

<system_warning>Token usage: 35000/200000; 165000 remaining</system_warning>

The system_warning tag name also appears in the Claude app’s list of classifier-triggered reminders; whether the app’s usage is this same context-budget message, a distinct injection sharing the tag, or both is unverified.

Models affected:

• Sonnet 4.5
• Haiku 4.5
• Sonnet 4.6
• Sonnet 5

For Opus-class and Mythos-class models, the docs point to the task budgets feature.

System reminders

(todo: document api-level system messages)

Not to be confused with mid-conversation system messages or the Claude app’s prompt injections.)

Safeguards and classifiers

Classifiers detect specific content in user requests and model responses. See this page for more info.