Claude Family: Constitutional AI in Practice

Anthropic's Claude models are decoder-only Transformers, distinguished from competitors mostly by post-training: every Claude model is aligned with Constitutional AI rather than purely RLHF, and they emphasize tool use and computer use.

Model	Context	Max output	Input/Output ($/1M)	Notable
Claude 3 Haiku	200K	4K	$0.25 / $1.25	Fast, cheap, ChatGPT-Free-tier alternative
Claude 3 Opus	200K	4K	$15 / $75	Highest-capability at launch
Claude 3.5 Sonnet (2024)	200K	8K	$3 / $15	Tool use, computer use
Claude 3.5 Haiku	200K	8K	$0.80 / $4	~65 tokens/s
Claude 3.7 Sonnet	200K	128K	$3 / $15	Extended thinking; GPQA 84.8%

The 128K max-output on Claude 3.7 Sonnet enables visible chain-of-thought reasoning — the model can produce up to 128K tokens of "thinking" before its final answer. This is comparable in shape to GPT-5's thinking mode and DeepSeek-R1.

Constitutional AI

Every Claude model is post-trained with Constitutional AI: a written set of principles guides the model's self-critique during fine-tuning. Compared to pure RLHF, this scales alignment without proportionally scaling human labeling, and the principles themselves are inspectable rather than implicit.

Code

Claude with extended thinking·python

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-3-7-sonnet-20250219",
    max_tokens=128_000,                  # extended-thinking budget
    thinking={"type": "enabled", "budget_tokens": 80_000},
    messages=[{"role": "user", "content": "Solve: integral of x*sin(x) dx"}]
)
print(response.content)
# response.content includes a 'thinking' block + a 'text' block.
# The model 'shows its work' inside thinking, then writes the final answer.

Claude Family: Constitutional AI in Practice

Constitutional AI

Code

External links

Exercise

Progress

Comments 0