Anthropic API¶

Knowing the Anthropic API signals you've worked with more than one provider and can reason about API design tradeoffs — a signal interviewers at companies that use Claude will explicitly look for.

Q1: What are the key differences between the OpenAI and Anthropic APIs?¶

Show answer

	OpenAI	Anthropic
System prompt	`{"role": "system", "content": "..."}` in messages	Separate `system` parameter
Streaming	Iterator of chunks	Context manager + event stream
Tool results	`{"role": "tool", "tool_call_id": "..."}`	`{"role": "user", "content": [{"type": "tool_result", ...}]}`
Tool arguments	String — must `json.loads()`	Already a dict — no parsing needed
Max tokens	`max_tokens` (optional, defaults vary)	`max_tokens` required
Model naming	`gpt-4o`, `gpt-4o-mini`	`claude-opus-4-7`, `claude-sonnet-4-6`, `claude-haiku-4-5`

The most common mistake when switching from OpenAI to Anthropic: trying to json.loads() tool arguments. Anthropic already parses them — block.input is a dict, not a string.

Q2: Write a basic Anthropic API call with a system prompt.¶

Show answer

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="You are a concise technical writer. Reply in plain text only.",
    messages=[
        {"role": "user", "content": "Explain what a transformer is in two sentences."}
    ],
)

print(response.content[0].text)

Key difference from OpenAI: max_tokens is required in Anthropic. The response content is a list — always access response.content[0].text for a standard text response.

Q3: How does Anthropic's streaming API work?¶

Show answer

Anthropic streaming uses a context manager and yields typed events rather than raw chunks.

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Write a haiku about embeddings."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

final = stream.get_final_message()
print(f"\nInput tokens: {final.usage.input_tokens}")
print(f"Output tokens: {final.usage.output_tokens}")

The context manager ensures the stream is properly closed. stream.text_stream yields text fragments only — it filters out non-text events. get_final_message() gives you the complete message with usage stats after the stream ends.

Q4: How does tool use work in the Anthropic API?¶

Show answer

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.getenv("ANTHROPIC_API_KEY"))

tools = [{
    "name": "get_weather",
    "description": "Get current weather for a city",
    "input_schema": {
        "type": "object",
        "properties": {"city": {"type": "string", "description": "City name"}},
        "required": ["city"],
    }
}]

messages = [{"role": "user", "content": "What's the weather in Tokyo?"}]
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=tools,
    messages=messages,
)

# Check for tool use
for block in response.content:
    if block.type == "tool_use":
        # block.input is ALREADY a dict — no json.loads() needed
        city = block.input["city"]
        weather_result = f"68°F, cloudy in {city}"

        # Append the assistant response and tool result
        messages.append({"role": "assistant", "content": response.content})
        messages.append({
            "role": "user",
            "content": [{
                "type": "tool_result",
                "tool_use_id": block.id,
                "content": weather_result,
            }]
        })

# Final call
final = client.messages.create(model="claude-sonnet-4-6", max_tokens=1024, tools=tools, messages=messages)
print(final.content[0].text)

The tool result goes back as a user message with type: "tool_result" — not a separate role like OpenAI's "tool" role.

Q5: Which Anthropic model should you use for which tasks?¶

Show answer

Model	Strengths	When to use
`claude-opus-4-7`	Most capable, best reasoning	Complex analysis, nuanced writing, hard coding tasks
`claude-sonnet-4-6`	Best balance of capability and speed	Most production workloads — RAG, agents, function calling
`claude-haiku-4-5`	Fastest and cheapest	Classification, routing, simple extraction, high-volume tasks

In a RAG system: use claude-haiku-4-5 for query analysis and routing, claude-sonnet-4-6 for answer generation. This splits cost while preserving quality at the generation stage.

Anthropic also supports prompt caching — if your system prompt is large (> 1024 tokens), you can cache it and pay only for read tokens on subsequent calls. Cache TTL is 5 minutes by default.

Previous: OpenAI API | Next: Embeddings