Anthropic API

Anthropic and Claude

Anthropic was founded by ex-OpenAI researchers focused on AI safety. Their Claude models are known for: exceptional instruction following, honest and harmless responses, excellent long-context understanding (200K tokens), and strong coding/analysis abilities.

Think of it like this: if OpenAI GPT is the Flipkart of AI (first mover, biggest market), Claude is the Amazon India - came later but with a focus on quality, reliability, and customer trust.

Claude Model Lineup (2025-26):

Where Claude Excels vs GPT-4:

• Instruction Following - Claude is exceptionally good at following complex, multi-step instructions precisely
• Long Context - 200K tokens natively. Can process entire codebases, long documents, books
• Honesty - More likely to say "I do not know" than hallucinate. Trained with Constitutional AI principles
• Safety - Better at refusing harmful requests while remaining helpful for legitimate use cases
• Code Quality - Claude Opus produces some of the best code among all models, especially for complex refactoring

Messages API Structure

The Anthropic Messages API uses a similar chat format to OpenAI but with some differences. The system prompt is a top-level parameter (not a message role), and responses use content blocks that can contain text, tool calls, or images.

Key Differences from OpenAI:

• System Prompt - Separate parameter, not in messages array. Anthropic recommends detailed system prompts for best results.
• Content Blocks - Response is an array of content blocks (text, tool_use, etc.) instead of a single content string.
• max_tokens is Required - You must explicitly set max_tokens. No default. This forces cost-conscious usage.
• No Implicit JSON Mode - Use tool_use with a schema or prompt engineering for structured output.
• Stop Reasons - end_turn (normal), max_tokens (truncated), stop_sequence (custom stop), tool_use (wants to call a tool).

Important Parameters:

• temperature - 0-1 range (vs OpenAI 0-2). Default 1.0. Use 0 for deterministic, 0.7 for creative.
• top_k - Unique to Anthropic. Limits token sampling to top K most likely tokens. Good for focused responses.
• stop_sequences - Custom strings that stop generation. Useful for structured output parsing.
• metadata.user_id - Pass user ID for abuse detection and rate limiting by Anthropic.

How Tool Use Works in Claude

Tool use is Anthropic version of function calling. You define tools with names, descriptions, and input schemas. Claude decides when to use them and returns a tool_use content block with structured input. You execute the tool and send back a tool_result message.

Tool Use Flow:

1. Define Tools - Provide tool definitions with JSON Schema in the API request
2. Claude Decides - Based on the conversation, Claude decides to call a tool. Response stop_reason = "tool_use"
3. Extract & Execute - Parse the tool_use block, execute the function in your code
4. Return Results - Send a user message with tool_result content block containing the function output
5. Claude Responds - Claude incorporates the result and generates the final response

Tool Use Best Practices:

• Detailed Descriptions - Claude reads tool descriptions carefully. More detail = better tool selection
• Example Values - Include examples in parameter descriptions for better accuracy
• Tool Choice Control - Use tool_choice: "auto" (Claude decides), "any" (must use some tool), or specific tool name
• Multiple Tools Per Turn - Claude can request multiple tool calls in one response
• Error Handling - Return clear error messages in tool_result. Claude can reason about errors and try alternatives

Computer Use (Beta):

Claude can control a computer - move mouse, click buttons, type text, take screenshots. This enables automation of any GUI-based workflow. Currently in beta but revolutionary for tasks like filling forms, navigating websites, and testing UIs.

Streaming with Server-Sent Events

Anthropic uses SSE (Server-Sent Events) for streaming. Events include: message_start, content_block_start, content_block_delta (actual text chunks), content_block_stop, and message_stop. This gives you fine-grained control over the streaming lifecycle.

• Token-by-Token - Each delta contains one or more tokens for real-time UI updates
• Tool Use Streaming - Tool call inputs are also streamed, so you can start processing before the full input is received

Prompt Caching:

Anthropic offers prompt caching - mark parts of your prompt as cacheable, and if the same content appears in subsequent requests, you get a 90% discount on those input tokens. The cache has a 5-minute TTL.

• Cache system prompts - Your 2000-token system prompt costs full price once, then 90% off for 5 minutes
• Cache tool definitions - Tool schemas stay the same across requests
• Cache long documents - RAG context that multiple users query

Extended Thinking:

Claude can be asked to "think" before responding, showing its chain-of-thought reasoning. This is enabled via the thinking parameter and is especially useful for complex problems.

• Thinking tokens are visible to you but can be hidden from users
• Dramatically improves accuracy on math, logic, and multi-step reasoning
• You pay for thinking tokens but the quality improvement is significant

Batches API:

Like OpenAI Batch API - submit large batches of requests and get results within 24 hours at 50% discount. Perfect for bulk processing, evaluation runs, and data labeling.

Pattern 1: Long Document Analysis

Claude 200K context window is perfect for analyzing entire codebases, legal contracts, or research papers in a single request. No need for chunking or RAG for documents under 200K tokens (~150K words).

• Upload the full document as a user message
• Ask specific questions or request analysis
• Claude can cross-reference information across the entire document

Pattern 2: Multi-step Agent with Tool Use

Build an agent that uses Claude tool use for multi-step tasks. Claude excels at knowing when to use tools vs when to respond directly, reducing unnecessary tool calls.

Pattern 3: Tiered Model Strategy

• Haiku - Classification, routing, simple extraction. Fast and cheap.
• Sonnet - Most tasks: chat, coding, analysis, writing. Best value.
• Opus - Complex reasoning, research, novel problem solving. When quality is everything.

Common Mistakes:

• Not setting max_tokens - Required parameter. Forgetting it causes API errors.
• Vague system prompts - Claude responds much better to detailed, specific instructions.
• Ignoring prompt caching - Leaving money on the table if you have repeated system prompts.
• Not using Haiku for simple tasks - Using Opus for classification is like using a bulldozer to plant a flower.

Q: What are the key differences between Anthropic API and OpenAI API?

Q: How does tool use work in Claude and what makes it reliable?

Q: What is prompt caching in Anthropic API and when would you use it?

Q: When would you choose Claude over GPT-4 for a project?

Q: How would you design a cost-effective multi-model strategy using Claude models?