Constitutional AI: how Anthropic builds safety in

Claude is trained using Constitutional AI (CAI), a technique Anthropic invented that supplements human preference labeling with a written 'constitution' — a set of principles the model uses to critique and revise its own outputs. This lets Anthropic scale alignment without requiring humans to hand-label millions of harmful examples, and it makes Claude's refusals more principled: it can usually explain why it will not do something rather than refusing blankly.

• Step 1 (SL-CAI): the model answers red-team prompts, then critiques and rewrites its own responses against the constitution
• Step 2 (RL-CAI): a preference model is trained on those AI-generated critiques and used as the reward signal for reinforcement learning
• The constitution draws on sources such as the UN Declaration of Human Rights plus Anthropic's own principles
• Result: more consistent, explainable alignment that scales without armies of human labelers

CAI vs standard RLHF: Standard RLHF relies entirely on human raters to mark what is harmful. Constitutional AI adds model self-critique against explicit principles, which makes alignment more consistent and far easier to scale.

The Claude model family: from Claude 3 to the 5.5 generation

Anthropic uses three named tiers within each generation — Haiku (fast and cheap), Sonnet (balanced default), and Opus (maximum capability) — plus, as of June 2026, a Mythos-class tier above Opus. A consistent pattern holds across generations: each new Sonnet tends to match or beat the previous generation's Opus at a lower price.

Where the line is now (November 2026): Claude 3.5 Sonnet is historic, not current. Anthropic's list prices are Fable 5.1 at $10 / $50, Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10, and Haiku 5.5 from $0.10 / $0.50 per million input/output tokens, all with a 1M-token context window and 128K maximum output. Opus 4.8, Sonnet 4.6, and Haiku 4.5 are now legacy models that remain available. Fable 5 was pulled under a US export-control directive in June 2026; the controls were lifted on June 30 and access returned on July 1.

Extended and adaptive thinking

Claude 3.7 Sonnet introduced 'extended thinking' — the model reasons step by step inside a separate token budget before producing its visible answer, similar in spirit to OpenAI's o-series. In the 4.x generation this evolved into adaptive thinking: instead of toggling a mode, you ask for harder reasoning and the model scales its own depth, spending more internal compute only when the task warrants it.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user",
               "content": "Prove that the square root of 2 is irrational."}],
)

for block in response.content:
    if block.type == "thinking":
        print("REASONING:", block.thinking[:200])
    elif block.type == "text":
        print("ANSWER:", block.text)

This is the legacy pattern: The example above uses manual extended thinking on Sonnet 4.6. Anthropic's docs mark thinking.type: "enabled" with budget_tokens as deprecated on Opus 4.6 and Sonnet 4.6 and not accepted on later models. Current models use adaptive thinking, steered by the effort setting.

What Claude is best at

• Agentic coding and large-codebase work — Claude Code is one of the most widely used AI coding tools
• Long-document analysis using the 1M-token context window
• Careful instruction following and brand-safe, nuanced writing
• Safety-sensitive deployments where consistent, explainable refusals matter

Anthropic suggests starting with Opus 5.5 for most workloads. Fable 5.1 is for the most demanding reasoning and long-horizon agent work, or when evals on Opus 5.5 still fall short. Sonnet 5.5 costs half as much as Opus 5.5, and Haiku 5.5 handles high-volume, latency-sensitive traffic. A routing strategy across tiers usually beats defaulting everything to the most capable model.

Benchmarks and key specs

How fast this moved: Claude 3.5 Sonnet set a coding-benchmark record at launch in 2024 with about 49% on SWE-bench Verified. Less than two years later, Sonnet 4.6 reaches ~79.6% and Opus 4.8 reaches 88.6% — a vivid illustration of how quickly the frontier moved. Numbers are vendor-reported, and a dash means this table lists no verified score for that model.

Frequently asked questions

Is Claude 3.5 Sonnet still the latest? No. It was a milestone in 2024, but the current family is Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5.

What is Constitutional AI? It is Anthropic's alignment method: the model critiques and revises its own outputs against a written set of principles (a 'constitution'), which scales alignment without relying solely on human labels and makes refusals more principled.

Which Claude model should I use? Anthropic suggests Opus 5.5 for most workloads. Use Fable 5.1 for the hardest long-horizon work, Sonnet 5.5 when speed and cost matter more, and Haiku 5.5 for high-volume, latency-sensitive tasks.

Does Claude have a free tier? Yes — claude.ai offers free access with limits, while paid Pro and Max plans raise usage caps and unlock the most capable models.