LLM API Cheatsheet#

All examples here are verified against the canonical vendor docs as of 2026-04-10. If you are using LangChain, prefer from langchain.chat_models import init_chat_model with a provider-prefix string ("anthropic:claude-sonnet-4-6") instead of these raw SDKs — see LangGraph Foundations & State Management.

Anthropic (Claude)#

Install#

pip install anthropic

Python 3.9+ required. Set ANTHROPIC_API_KEY.

Basic message#

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(message.content[0].text)

Current model IDs#

Model

ID

Context

Max out

Cost (in/out per MTok)

Opus 4.6

claude-opus-4-6

1M

128k

\(5 / \)25

Sonnet 4.6

claude-sonnet-4-6

1M

64k

\(3 / \)15

Haiku 4.5

claude-haiku-4-5

200k

64k

\(1 / \)5

Tool use (client tools)#

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=[{
        "name": "get_weather",
        "description": "Get current weather for a city",
        "input_schema": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
        "strict": True,
    }],
    messages=[{"role": "user", "content": "Weather in SF?"}],
)
# response.content contains tool_use blocks — execute and send back tool_result

Server tools (Anthropic runs them)#

Available types: web_search_20260209, code_execution, web_fetch, tool_search.

response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=1024,
    tools=[{"type": "web_search_20260209", "name": "web_search"}],
    messages=[{"role": "user", "content": "Latest on Mars rover?"}],
)

Extended thinking#

# Opus 4.6 / Sonnet 4.6 — adaptive thinking (budget_tokens is deprecated)
client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    messages=[{"role": "user", "content": "Is there an infinite number of primes p with p mod 4 == 3?"}],
)

# Older models — explicit budget
client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[...],
)

Streaming#

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

OpenAI#

Install#

pip install openai

Set OPENAI_API_KEY.

Responses API (primary in current docs)#

from openai import OpenAI
client = OpenAI()

response = client.responses.create(
    model="gpt-5.2",
    input="Why do parrots talk?",
)
print(response.output_text)

Chat Completions (legacy but supported)#

completion = client.chat.completions.create(
    model="gpt-5.2",
    messages=[{"role": "user", "content": "Hello"}],
)
print(completion.choices[0].message.content)

Tool use (Chat Completions shape)#

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

response = client.chat.completions.create(
    model="gpt-5.2",
    messages=[{"role": "user", "content": "Weather in SF?"}],
    tools=tools,
    tool_choice="auto",
)

Streaming#

stream = client.chat.completions.create(
    model="gpt-5.2",
    messages=[{"role": "user", "content": "Count to 5"}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta.content or ""
    print(delta, end="", flush=True)

Google Gemini#

Install#

pip install -q -U google-genai

Use the new google-genai package, not the older google-generativeai. Set GEMINI_API_KEY.

Basic generation#

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Explain attention in one sentence.",
    config=types.GenerateContentConfig(
        system_instruction="You are a concise ML teacher.",
        temperature=1.0,  # keep default for Gemini 3
    ),
)
print(response.text)

Streaming#

for chunk in client.models.generate_content_stream(
    model="gemini-3-flash-preview",
    contents="List 3 facts about penguins.",
):
    print(chunk.text, end="", flush=True)

Thinking#

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Solve: 23 * 47",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(thinking_budget=2048),
    ),
)

Which API should I use?#

  • Direct vendor SDKs — when you need a provider-specific feature (Anthropic extended thinking with server tools, OpenAI Responses API reasoning items, Gemini 3 multimodal chaining).

  • LangChain init_chat_model + create_agent — for everything else. Uniform interface, works with all three providers, and gives you streaming / persistence / human-in-the-loop for free.

See LangGraph Foundations & State Management for the LangChain/LangGraph path.

Practice#

1. Hello World in three providers#

Write a single Python file that calls claude-sonnet-4-6, gpt-5.2, and gemini-3-flash-preview with the same prompt ("Name one thing parrots and dolphins have in common.") and prints all three responses side by side. Each should live behind a function of the signature call(prompt: str) -> str.

Use environment variables for API keys; never hardcode.

2. Tool use round-trip#

Using the Anthropic SDK, define a get_weather(city) tool with strict: true. Implement the client-side loop:

  1. Send the user query.

  2. Parse response.content for a tool_use block.

  3. Execute the tool (return a stub "sunny, 72°F").

  4. Send the tool_result back to the same model.

  5. Print the final natural-language answer.

3. Streaming token counter#

Stream a ~500-token response from any provider and count the tokens as they arrive (simple word count is fine). Print the count every 50 tokens to demonstrate you are receiving partial chunks rather than waiting for the whole response.

4. Extended thinking budget comparison#

On Claude Sonnet 4.6, solve the same hard problem twice:

  1. With thinking={"type": "adaptive"}

  2. Without any thinking parameter

Compare:

  • Latency (wall-clock)

  • Quality (is the answer correct?)

  • Token usage (usage.output_tokens and any thinking token counts)

Write a one-paragraph takeaway about when extended thinking is worth the cost.

5. Provider migration#

Take a working Anthropic SDK script. Migrate it to use LangChain’s init_chat_model("anthropic:claude-sonnet-4-6") and confirm the output is equivalent. Then swap the model string to "openai:gpt-5.2" without changing any other code — confirm it still works. This demonstrates the value of the LangChain abstraction for multi-provider code.

Review Questions#

  1. What is the install command for the official Anthropic Python SDK?

    • A. pip install claude-api

    • B. pip install anthropic-sdk

    • C. pip install anthropic

    • D. pip install claude-python

  2. Which OpenAI API is documented as the primary surface for new code as of 2026?

    • A. Completions API (client.completions.create)

    • B. Chat Completions API (client.chat.completions.create)

    • C. Responses API (client.responses.create)

    • D. Assistants API

  3. Which package should you install for Google Gemini in Python today?

    • A. google-generativeai (the older package)

    • B. google-genai (the new SDK)

    • C. gemini-python

    • D. gcloud-ai

  4. On Claude Opus 4.6 and Sonnet 4.6, which thinking mode is recommended?

    • A. {"type": "enabled", "budget_tokens": 10000}

    • B. {"type": "adaptive"}

    • C. {"type": "disabled"}

    • D. Thinking is not supported

  5. What is the recommended temperature for Gemini 3 models?

    • A. 0.0

    • B. 0.5

    • C. 1.0 (default)

    • D. 2.0

  6. You need a 1M-token context window and top-tier reasoning. Which Claude model?

    • A. claude-haiku-4-5 (200k context)

    • B. claude-opus-4-6 (1M context)

    • C. claude-3-opus-20240229

    • D. claude-instant

  7. Which Anthropic server tool type enables the model to search the web?

    • A. web_search_20260209

    • B. google_search

    • C. web_tool

    • D. search_engine

  8. Which parameter guarantees an Anthropic tool call conforms to the declared input schema?

    • A. temperature: 0

    • B. strict: true on the tool definition

    • C. mode: "strict"

    • D. json_mode: true

  9. You want provider-agnostic code that can swap between Claude, GPT, and Gemini without rewrites. What do you reach for?

    • A. A custom HTTP wrapper

    • B. LangChain’s init_chat_model with a provider-prefix string

    • C. Three separate if branches per provider

    • D. openai package only

  10. Where should API keys come from in production code?

    • A. Hardcoded at the top of the file

    • B. Environment variables (or a secret manager)

    • C. A public GitHub gist

    • D. The user’s prompt

View Answer Key

  1. C — pip install anthropic.

  2. C — The Responses API is the primary surface in current OpenAI docs; Chat Completions is still supported as legacy.

  3. B — google-genai is the new SDK; google-generativeai is the older package.

  4. B — Adaptive thinking is the recommended mode for Opus/Sonnet 4.6; budget_tokens is deprecated on those models.

  5. C — Google explicitly recommends keeping temperature=1.0 default for Gemini 3.

  6. B — Claude Opus 4.6 has a 1M-token context and is the highest-intelligence broadly-available Claude.

  7. A — web_search_20260209 (dated identifiers are how Anthropic versions server tools).

  8. B — strict: true on the tool enforces schema conformance.

  9. B — init_chat_model("anthropic:claude-sonnet-4-6") vs "openai:gpt-5.2" — same code, different string.

  10. B — Environment variables or a secret manager. Never hardcode.