Anthropic Features

Anthropic API Features — March 2026 Reference

Master reference for all current Anthropic API capabilities. Use this skill when integrating Anthropic features into NodeJS-Starter-V1.

Model Registry

Implementation path: apps/backend/src/models/anthropic.py → ClaudeModels Selector: ModelSelector.get_client(tier="sonnet") → auto-selects by settings

Adaptive Thinking

Allows Claude to reason step-by-step before responding. Available on Opus 4.6 and Sonnet 4.6.

# Enable via create_message()
response = await client.create_message(
    messages=messages,
    thinking=True,
    thinking_budget=16000,  # 1024–128000
)

Setting: THINKING_ENABLED=true, THINKING_BUDGET_TOKENS=10000 Constraint: budget_tokens must be less than max_tokens (unless interleaved)

Fast Mode

2.5× faster output for Opus 4.6. Research preview — not for production latency SLAs.

response = await client.create_message(
    messages=messages,
    fast_mode=True,  # Opus 4.6 only
)

Setting: FAST_MODE_ENABLED=true Note: output_config: { speed: "fast" } — Opus 4.6 only

Automatic Prompt Caching (GA)

No beta header required. Caches system prompts and tool definitions across requests.

• System prompts > 2000 characters are auto-cached in create_message()
• Cache control: { "type": "ephemeral" } — 5-minute default TTL, optional "1h"
• Minimum cacheable: 1024 tokens

Implementation: AnthropicClient._create_cached_system() and _add_cache_to_tools()

Token Counting API

Free pre-flight call — no model inference, no charge.

token_count = await client.count_tokens(
    messages=messages,
    system=system_prompt,
)
# warn if token_count > settings.token_count_warning_threshold

Setting: TOKEN_COUNT_WARNING_THRESHOLD=50000 Use case: Cost estimation, input validation, context window management

Web Search Tool v2 (GA — Feb 2026)

Requires Opus 4.6 or Sonnet 4.6. Priced as tool-use tokens.

tool = AnthropicClient.get_web_search_tool(
    max_uses=5,
    allowed_domains=["docs.python.org"],   # optional
    blocked_domains=["reddit.com"],         # optional
)
response = await client.create_message(messages=messages, tools=[tool])

Type identifier: web_search_20260209 Setting: WEB_SEARCH_ENABLED=true, WEB_SEARCH_MAX_USES=5

Code Execution Tool (GA — Feb 2026)

Runs Python in a secure sandboxed environment. Free when bundled with web search.

tools = [
    AnthropicClient.get_web_search_tool(),
    AnthropicClient.get_code_execution_tool(),
]

Type identifier: code_execution_20250825

Agent Skills (Beta)

Anthropic-managed document processing. Beta header: skills-2025-10-02. Supported: excel, powerpoint, word, pdf.

skill = AnthropicClient.get_agent_skill("pdf")

Setting: AGENT_SKILLS_ENABLED=true

MCP Connector (Beta) — "Remote Control"

Connect any MCP-compatible remote server in the messages API. Beta header: mcp-connector-2025-05-14.

server = AnthropicClient.get_mcp_server_tool(
    server_name="github",
    server_url="https://mcp.github.com",
    allowed_tools=["create_issue", "list_repos"],
)

Setting: MCP_CONNECTOR_ENABLED=true

Structured Outputs

Enforced JSON schema response — model guaranteed to match schema.

schema = {
    "name": "task_result",
    "schema": {
        "type": "object",
        "properties": {"status": {"type": "string"}, "result": {"type": "string"}},
        "required": ["status", "result"],
    },
}
response = await client.create_message(messages=messages, schema=schema)

Voice Mode

NOT available via Anthropic API. Voice is only available in:

• Claude.ai consumer app
• Claude Code CLI

There is no voice API endpoint. Do not implement or promise voice features via API.

Out of Scope (Deferred)

• Message Batches API — 50% cost reduction for async bulk jobs. Requires job queue.
• Files API — Upload-once file references. Requires file management UI.