Claude Model Family Explained: Opus, Sonnet, and Haiku

One of the first decisions you make when working with the Anthropic API is which model to use. Anthropic does not offer a single all-purpose model — it offers a family of models, each optimised for a different balance of intelligence, speed, and cost.

Choosing the wrong model does not just affect quality. It affects how much you pay, how fast your application responds, and whether your users get the experience they expect. Choosing the right one from the start saves time and money.

This guide breaks down the entire Claude model family as of 2026 — what each model is designed for, how they compare on real tasks, and how to decide which one belongs in your application.

Why a Model Family Instead of One Model?

This is a question worth asking. Why does Anthropic build three models instead of one perfect model for everything?

The answer comes down to a fundamental trade-off in AI development: intelligence vs. efficiency. More capable models require more computation — more memory, more processing time, higher energy cost — to produce their outputs. For some tasks, that extra computation produces dramatically better results. For others, a faster, cheaper model produces results that are just as good.

A model that is overkill for a task wastes money and slows your application. A model that is underpowered for a task produces poor results that undermine your product. The model family lets you match the tool to the task.

The Naming Convention

Claude models follow a consistent naming pattern: the model name (Opus, Sonnet, Haiku) plus a generation number (4, 4.5, 4.6) and optionally a snapshot date. For example, claude-opus-4-6 and claude-haiku-4-5-20251001. When you use an alias like claude-sonnet-4-6 without a date, you get Anthropic's recommended version of that model — which may be updated over time.

Claude Opus 4.6 — Maximum Intelligence

Claude Opus 4.6 is the most capable model in the current Claude family. It is Anthropic's answer to the question: what happens when you push the boundaries of what a language model can do?

What Makes Opus Different

Opus is trained and optimised to excel at tasks that require sustained, multi-step reasoning. It scores highest on the most demanding academic and professional benchmarks:

GPQA: Graduate-level questions in physics, chemistry, and biology requiring genuine expert reasoning
SWE-bench: Real software engineering tasks from GitHub repositories — not toy coding problems
MATH: Competition-level mathematics requiring multi-step derivations
HumanEval: Code generation tasks evaluated by functional correctness

Extended Thinking with Opus

Opus 4.6 supports adaptive thinking — the ability to dynamically decide how much reasoning to apply before giving an answer. For complex problems, Opus will work through a series of internal reasoning steps before producing its response, similar to how a human expert might think through a difficult problem before speaking.

You can control this behaviour with the effort parameter: setting effort high tells Claude to think harder; setting it low produces faster, more direct responses.

Context Window and Output

Context window: 1 million tokens
Max output: 128,000 tokens (synchronous API), up to 300,000 tokens via the Batch API with the extended output beta header
Knowledge cutoff: August 2025

When to Use Opus

Complex research tasks requiring synthesis of large volumes of information
Advanced coding workflows — architecture design, debugging complex systems, full codebase understanding
Multi-step agentic tasks where the model must plan, execute, and recover from errors autonomously
High-stakes professional work in legal, medical, or financial domains where accuracy is critical
Tasks where you are processing very long documents and need strong attention to detail throughout

Opus Pricing

Input: $5 per million tokens
Output: $25 per million tokens

Opus is an Investment

Opus is the most expensive model in the family. Use it for tasks where the quality difference genuinely matters — complex analysis, critical decisions, difficult reasoning. For anything routine, Sonnet gives you most of the benefit at a third of the cost.

Claude Sonnet 4.6 — The Balanced Workhorse

Claude Sonnet 4.6 is the model that most developers and organisations should start with. It delivers near-Opus intelligence at significantly lower cost and faster response times.

What Makes Sonnet the Default Choice

Sonnet sits at the sweet spot of the capability-cost curve. On most practical professional tasks — summarising documents, generating code, answering complex questions, running customer support interactions, content generation — Sonnet performs at a level that is indistinguishable from Opus to most users, while costing 40% less on input and 40% less on output.

It is fast enough for interactive applications where users are waiting for a response, yet powerful enough for nuanced, long-form work.

Sonnet's Capabilities

Extended thinking: Yes — Sonnet supports reasoning modes for complex tasks
Context window: 1 million tokens — identical to Opus
Max output: 64,000 tokens (synchronous), up to 300,000 via Batch API
Knowledge cutoff: January 2026
Vision: Full image and document analysis
Tool use: All tools including web search, code execution, and custom client tools

When to Use Sonnet

Production APIs that serve real users — you need speed and reliability at scale
Customer support, document processing, and content pipelines
Coding assistance, code review, and automated testing
RAG (retrieval-augmented generation) applications where Claude ingests retrieved context
Anything you are scaling to significant usage volume and need cost control

Sonnet Pricing

Input: $3 per million tokens
Output: $15 per million tokens

Claude Haiku 4.5 — Speed and Efficiency at Scale

Claude Haiku 4.5 is the smallest and fastest model in the family. It is designed for applications where response time is the primary constraint and where tasks are well-defined enough that a smaller model can handle them reliably.

What Makes Haiku Unique

Haiku's defining characteristic is its speed. It produces responses significantly faster than Sonnet or Opus, making it suitable for real-time interfaces where even a two-second delay feels like lag. It is also the most cost-efficient model in the family by a wide margin.

Haiku does not have the reasoning depth of Opus or Sonnet. But for tasks that are well-scoped and do not require complex multi-step thinking, it performs extremely well.

Haiku's Capabilities

Extended thinking: Not supported
Context window: 200,000 tokens
Max output: 64,000 tokens
Knowledge cutoff: February 2025
Vision: Yes — full image analysis
Tool use: Full tool use support

When to Use Haiku

Real-time chat interfaces where latency is the top priority
High-volume classification, routing, or tagging tasks
Simple question-answering where the query is well-defined
Pre-processing steps in a larger pipeline — extract key information before passing to Sonnet or Opus for deeper analysis
Cost-sensitive applications where you are processing millions of tokens per day

Haiku Pricing

Input: $1 per million tokens
Output: $5 per million tokens

Haiku as a First Pass

A common production pattern is to use Haiku as a classifier or triage model — quickly determining whether a request is simple (answer with Haiku) or complex (escalate to Sonnet or Opus). This hybrid approach dramatically reduces average cost per request without sacrificing quality on complex tasks.

Side-by-Side Comparison

Here is a structured comparison of the three models across the dimensions that matter most for developers making a choice:

Intelligence level: Opus (highest) → Sonnet (near-frontier) → Haiku (near-frontier, narrower tasks)
Context window: Opus 1M tokens | Sonnet 1M tokens | Haiku 200K tokens
Max output: Opus 128K | Sonnet 64K | Haiku 64K
Extended thinking: Opus ✓ | Sonnet ✓ | Haiku ✗
Latency: Opus (moderate) | Sonnet (fast) | Haiku (fastest)
Input cost per MTok: Opus $5 | Sonnet $3 | Haiku $1
Output cost per MTok: Opus $25 | Sonnet $15 | Haiku $5
Knowledge cutoff: Opus Aug 2025 | Sonnet Jan 2026 | Haiku Feb 2025

Available Access Methods

All three models are accessible through multiple platforms, giving you flexibility in how you deploy and pay for them.

Direct Anthropic API

The primary access method is through api.anthropic.com. This gives you the latest model versions first, direct billing with Anthropic, and access to all features including beta capabilities.

Amazon Bedrock

Claude is available through AWS Bedrock using model IDs like anthropic.claude-opus-4-6-v1. This is ideal for organisations already operating in AWS, as it integrates with AWS IAM, CloudWatch, and consolidated billing.

Google Cloud Vertex AI

Claude is available on GCP Vertex AI using identifiers like claude-opus-4-6. Ideal for organisations running Google Cloud infrastructure.

Microsoft Azure AI Foundry

Claude is available in preview on Microsoft Azure, with regional deployment options and Azure Active Directory integration.

Feature Availability Varies by Platform

Not all Claude features are available on every platform. Extended thinking, the Files API, MCP connector, and some beta features are available on the direct Anthropic API first, and may arrive on Bedrock and Vertex AI later. If cutting-edge features matter for your application, the direct API is the best starting point.

How to Choose: A Decision Framework

If you are unsure which model to start with, follow this simple decision process:

Is your task complex, open-ended, or high-stakes? Examples: analysing a 200-page contract, building an autonomous coding agent, synthesising research across many documents. Start with Opus 4.6.
Is your task moderately complex and production-facing? Examples: customer support, document summarisation, code review, content generation. Start with Sonnet 4.6.
Is your task simple, well-defined, or time-critical? Examples: real-time chat, classification, simple Q&A, high-volume batch processing. Start with Haiku 4.5.
Are you unsure? Start with Sonnet 4.6. It is the best default for discovering what your workload actually needs before optimising.

Summary

Anthropic's three-tier model family — Opus, Sonnet, and Haiku — gives you the tools to build applications that are intelligent where intelligence matters and efficient where efficiency matters. Understanding when to use each model is one of the core skills you develop as a Claude developer.

As you build more applications and run more experiments, you will develop intuition for which model fits which workload. The next step is actually getting access and making your first API call.

In our next post, we step away from the API and start with the consumer-facing product: Getting Started with Claude.ai: Your First Conversation.

This post is part of the Anthropic AI Tutorial Series. Don't forget to check out our previous post: Claude vs ChatGPT vs Gemini: Which AI Should You Use in 2026?.