# How do Claude 3 Opus, Sonnet and Haiku differ?

Date: 2024-03-15 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/claude-3-opus-sonnet-haiku-differ/

**Answer:** The three Claude 3 models share the same basic abilities, including reading images and a 200,000-token context window, but trade intelligence against speed and price. Opus is the most capable and the most expensive, Sonnet balances quality and speed at a fifth of Opus's price, and Haiku, released this week, is the fastest and by far the cheapest.

When Anthropic announced Claude 3 on March 4, it released two of the three models right away and promised the smallest soon. On March 13, Claude 3 Haiku arrived in the API and on claude.ai for Pro subscribers, so the whole family can now be compared side by side. The question for most people is no longer "what is Claude 3?" but "which one should I point at this job?"

Anthropic's own framing is a ladder: Haiku, Sonnet and Opus, in rising order of capability. Each step up is smarter and slower and costs more. Choosing well means matching the step to the task instead of always reaching for the top.

## The three models in one table

Prices are for the API and are counted in tokens, the pieces of words that models read and write. Input is what you send; output is what the model writes back.

| | Haiku | Sonnet | Opus |
|---|---|---|---|
| Position | Fastest, most compact | Balance of intelligence and speed | Most intelligent |
| Price per million tokens (input / output) | $0.25 / $1.25 | $3 / $15 | $15 / $75 |
| Speed, per Anthropic | About 21,000 tokens (roughly 30 pages) read per second on prompts under 32,000 tokens | About twice as fast as Claude 2 and 2.1 for most work | About the same speed as Claude 2 and 2.1 |
| Where you can use it | API, claude.ai Pro, Amazon Bedrock | API, free claude.ai, Amazon Bedrock | API, claude.ai Pro |
| Anthropic's suggested uses | Live customer support, translation, content moderation, pulling facts from messy data | Search over large knowledge bases, product recommendations, code generation, reading text in images | Complex multi-step tasks, research review and brainstorming, analysis of charts and financials |

All three accept images such as photos, charts and diagrams, and all start with a 200,000-token context window, the amount of text they can consider at once.

## What the price gap means in practice

The price difference is large: Opus costs 60 times as much as Haiku per token. A worked example shows what that means for one job. Say you summarize a report of about 10,000 tokens (around 7,500 words) into a 500-token summary.

- **Haiku:** 10,000 × $0.25 per million + 500 × $1.25 per million ≈ $0.003
- **Sonnet:** 10,000 × $3 per million + 500 × $15 per million ≈ $0.04
- **Opus:** 10,000 × $15 per million + 500 × $75 per million ≈ $0.19

For one report, none of these matters. For 10,000 reports a month, it is roughly $30 against $1,900. That is why Anthropic pitches Haiku at companies processing large volumes of documents, and why Opus makes sense when a single answer has to be as good as possible.

On quality, Anthropic reports that Opus beats OpenAI's GPT-4 and Google's Gemini 1.0 Ultra on several common benchmarks, including graduate-level reasoning and grade-school math. As Tom's Guide pointed out, that comparison is with a GPT-4 model that is a year old. Benchmarks are a starting point; your own tasks are the real test.

## A three-task test you can run

The best way to choose is to give all three models the same small set of tasks that look like your real work. Claude Pro subscribers can use all three models on claude.ai, and developers can call any of them through the API. Here is a set that separates them well:

```prompt
Task 1 (quick): Rewrite this customer email as a polite two-sentence reply
that confirms the refund and gives no new promises: [paste email]

Task 2 (structured): From the text below, extract every date, amount and
person's name into a table with three columns. [paste a page of notes]

Task 3 (hard): Here are two short proposals. Compare their risks, say which
assumptions are weakest in each, and recommend one in 150 words.
[paste both]
```

Score each answer from 1 to 5 on accuracy and usefulness, and note how long you waited. In a typical run, which is illustrative, all three handle the first task well, Haiku is clearly fastest on the first two, and the gap in quality only shows on the third, where Opus is more likely to catch a weak assumption.

> **Try it:** Run the three tasks above with your own material on two of the models. If the cheaper model scores the same on a task, you have found a job you don't need to pay more for.

## Which should you use for which task?

Use **Haiku** for high-volume, well-defined work where speed matters: answering routine questions in a chat window, sorting or tagging messages, translating short texts, and pulling structured data out of documents or images. It is also the sensible choice when you are building something that will run thousands of times a day.

Use **Sonnet** as your everyday default. It powers the free version of claude.ai, it is quick, and it handles drafting, summarizing, coding help and questions about long documents well. If you are not sure which model to start with, start here.

Use **Opus** when the task is open-ended, has many steps, or the cost of a weak answer is high: analyzing a contract or a dense research paper, planning a complex project, or reasoning through a tricky decision. If you subscribe to Claude Pro, a good habit is to draft with Sonnet and switch to Opus only for the parts where you need its extra judgment.

## Sources

- Introducing the next generation of Claude: https://www.anthropic.com/news/claude-3-family
- Claude 3 Haiku: our fastest model yet: https://www.anthropic.com/news/claude-3-haiku
- Claude 3 launched by Anthropic — new AI model leaves OpenAI's GPT-4 in the dust: https://www.tomsguide.com/ai/claude-3-launched-by-anthropic-new-ai-model-leaves-openais-gpt-4-in-the-dust