How Claude Works: Constitutional AI, 200K Context, and What Makes It Different

Anthropic's Constitutional AI training approach, the Opus/Sonnet/Haiku family, extended thinking, and where Claude outperforms competing frontier models.

Claude is Anthropic's AI assistant — and the model underneath it was built differently from other frontier models. Not just different architecture choices, but a fundamentally different philosophy about how to align an AI system with human values. That difference shows up in how Claude behaves at the edges: when asked something ambiguous, something borderline, or something where the right answer is genuinely hard.

[Video: Andrej Karpathy — Intro to Large Language Models (the best 1-hour overview of how frontier models like Claude are built from pretraining through fine-tuning)]

The model family (as of 2025)

Constitutional AI: how Claude is aligned

Most AI assistants are aligned via RLHF — human labellers rank model outputs, a reward model learns human preferences, and the model is trained via RL to maximise that reward. Anthropic's Claude uses a different approach: Constitutional AI (CAI).

In CAI, the model is given a set of explicit principles — the 'constitution' — and trained to critique and revise its own outputs against those principles. Instead of relying entirely on human labellers to identify bad outputs, the model learns to reason about whether its outputs satisfy the principles directly.

• Phase 1 — Supervised: model critiques its own harmful outputs and generates revisions; trains on those revisions
• Phase 2 — RLAIF: a feedback model (trained on the constitution) scores outputs instead of human labellers; RL on those scores
• Result: a model that can reason about its own values, not just pattern-match to previously-labelled harmful outputs

200K context window — and the lost-in-the-middle caveat

200K tokens (~150,000 words, ~500 pages) is a genuine capability advantage. You can pass an entire codebase, a full legal agreement, or a large corpus of documents in a single call. But there's a catch: attention doesn't distribute evenly across a long context.

Research (and empirical testing) consistently shows that information in the middle of a very long context is attended to less reliably than information near the start or end. For retrieval tasks in long-context RAG, this means your most relevant passage should be at the top or bottom of the injected context — not buried in the middle.

Prompt caching — Claude's cost advantage at scale

Claude's API supports prompt caching: if you mark a prefix of your prompt as cacheable, subsequent requests that share that prefix pay only 10% of the input token cost for the cached portion. For applications with long system prompts or large shared context (documentation, policy docs, codebase), this can reduce input costs by 70–90%.

import anthropic

client = anthropic.Anthropic()

# Cache a long system prompt — pay full cost first time, 10% on subsequent calls
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": open("large_system_prompt.txt").read(),  # e.g. 10K tokens
            "cache_control": {"type": "ephemeral"}           # mark for caching
        }
    ],
    messages=[{"role": "user", "content": "Summarise the key risks."}]
)

# On subsequent calls with the same system prompt prefix:
# First 10K tokens cost 10% of normal input price — cache hit
print(response.usage.cache_read_input_tokens)   # tokens read from cache
print(response.usage.cache_creation_input_tokens)  # tokens written to cache

Extended thinking — Claude's reasoning mode

Claude's extended thinking feature allows the model to generate an internal chain of reasoning (visible in the API response as thinking blocks) before producing its final answer. This is most useful for: complex multi-step reasoning, hard math or coding problems, and tasks where explicit step-by-step planning improves quality.

What Claude is distinctively good at vs. peers

The Anthropic API — quick start

import anthropic

client = anthropic.Anthropic()  # uses ANTHROPIC_API_KEY env var

# Basic generation
msg = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="You are an expert in production AI systems.",
    messages=[{"role": "user", "content": "What is prompt caching?"}]
)
print(msg.content[0].text)

# With tool use
tools = [{
    "name": "get_weather",
    "description": "Get current weather for a location",
    "input_schema": {
        "type": "object",
        "properties": {"location": {"type": "string"}},
        "required": ["location"]
    }
}]

msg = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=tools,
    messages=[{"role": "user", "content": "What's the weather in London?"}]
)
# If stop_reason == "tool_use", extract tool call and execute

Compare Claude models in Playground →: Run the same prompt across Claude Haiku, Sonnet, and Opus and see quality and speed differences live.