The Model Family: Opus, Sonnet and Haiku

Behind every Claude conversation is a specific model, and on paid plans you choose which one. Anthropic organises its models into three tiers with persistent names: Opus, Sonnet and Haiku. Version numbers after the name change as new generations ship, and a given tier's exact capabilities shift with each release, so the exam tests the principle of the tiers rather than any particular version. For the current line-up, context sizes and availability, consult the model overview in Anthropic's documentation.

The three tiers

Opus is the most capable tier. It is the choice for complex reasoning, nuanced judgement, long multi-step tasks and work where a wrong answer is expensive. It is also the slowest and, for API users, the most costly, which is why it is not the default for everything.

Sonnet is the balanced tier: strong intelligence at good speed. For most everyday knowledge work (drafting, summarising, analysis of ordinary complexity, coding) Sonnet is the sensible default, and it is typically the model the apps select unless you choose otherwise.

Haiku is the fastest and lightest tier. It handles straightforward tasks quickly: classification, extraction, short answers, simple reformatting, high-volume repetitive work. It trades some depth for speed and efficiency.

A helpful mental model is a team: Haiku is the quick assistant, Sonnet the capable generalist, Opus the senior specialist. You would not send every question to the senior specialist, and you would not ask the quick assistant to write the strategy.

Choosing a model in practice

Ask three questions.

1. How hard is the task? Ambiguous judgement, long documents with many interacting constraints, or subtle writing favour Opus. Clear tasks with clear criteria suit Sonnet. Mechanical tasks suit Haiku.
2. What does an error cost? The higher the stakes, the stronger the case for the top tier plus thorough verification.
3. How much do speed and volume matter? If you are processing dozens of similar items or want near-instant responses, favour the lighter tiers.

Usage limits on consumer plans are affected by model choice; heavier models consume more of your allowance per message. Using Sonnet or Haiku for routine work leaves more headroom for Opus when you need it.

In the apps, the model picker sits near the message box. Free plan users typically have access to a default model only; paid plans unlock the full selection. You can change model mid-conversation, which is a useful pattern: draft with Sonnet, then switch to Opus for a final critical review.

Extended thinking

Extended thinking is a mode in which Claude reasons at length before answering, working through the problem step by step. It noticeably improves results on maths, logic, planning, complex analysis and tasks with many constraints, at the cost of slower responses. In the apps it is a toggle; the exact name and placement vary by surface. Turn it on for hard problems where you want careful reasoning; leave it off for quick drafts, chat and simple lookups where it adds delay without benefit.

Extended thinking is not a substitute for verification. A model that has reasoned carefully can still start from a wrong fact.

The context window

The context window is the amount of text a model can hold in view at once: your messages, Claude's replies, attached files and any Project knowledge in use, all measured in tokens (roughly three-quarters of a word each in English). Every Claude model has a large window, large enough for several long reports at once, and some models and plans offer larger windows still. The practical implications:

• Very long conversations eventually approach the limit; the app will warn you and you may need to start a fresh chat.
• A huge document may not fit alongside a long conversation; attach it to a fresh chat or use a Project.
• Everything in the window influences the answer, including instructions you gave fifty messages ago.

Exact window sizes are published in the documentation and differ by model, so do not memorise a number; remember the concept and where to look it up.

Knowledge cut-off

Each model was trained on data up to a certain date. It does not know about events after that date unless you tell it or it uses web search. When a task depends on recent information (a new regulation, this quarter's results, a competitor's launch) enable search or supply the material. Asking Claude "what is the latest version of X?" without search is an invitation to a confident, outdated answer.

Scenario: choosing for a day's work

A policy analyst has three tasks. Tagging 150 consultation responses by theme: Haiku, fast and mechanical, with a sample of tags checked by hand. Drafting a summary briefing for the minister: Sonnet, with a clear rubric. Reviewing the final briefing for logical gaps and weak arguments: Opus with extended thinking, because this is the step where depth pays and where a missed flaw would be costly.

Exam focus

• Opus is the most capable tier for complex, high-stakes work; Sonnet is the balanced default; Haiku is the fastest tier for straightforward, high-volume tasks.
• Choose a model by task difficulty, cost of error and need for speed or volume; heavier models consume more of a plan's usage allowance.
• Extended thinking improves reasoning on hard problems at the cost of speed; it does not replace fact verification.
• The context window is the total text in view, measured in tokens; long chats and large files can approach it, prompting a fresh conversation or a Project.
• Models have a knowledge cut-off; for recent facts, enable web search or supply the information.