# What are Claude 3.7 Sonnet and Claude Code, and what can they do?

Date: 2025-02-25 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-3-7-sonnet-and-claude-code/

**Answer:** Claude 3.7 Sonnet, released by Anthropic on February 24, 2025, is a single model that can either reply right away or think through a problem step by step first, with that thinking shown to you. Alongside it, Anthropic opened a limited preview of Claude Code, a tool that lets Claude read, edit, test and commit code from a programmer's terminal.

Yesterday, Monday, February 24, 2025, Anthropic released **Claude 3.7 Sonnet**, the newest version of its mid-sized Claude model, and a second product aimed squarely at programmers, **Claude Code**. The model is the headline for most people: it is Anthropic's first model that can "reason", meaning it can spend extra time working through a problem before it writes the answer you see.

What makes it unusual is that it does this inside one model. Other companies offer a fast chat model and a separate, slower reasoning model, and you have to pick between them. Anthropic calls Claude 3.7 Sonnet a **hybrid reasoning model**: you decide, question by question, whether it answers immediately or thinks first. (If you are wondering about the number, Anthropic skipped 3.6; the previous version was Claude 3.5 Sonnet.)

## What Anthropic released

The model is available on every Claude plan, Free, Pro, Team and Enterprise, and to developers through Anthropic's API (the service programs use to send requests to Claude), Amazon Bedrock and Google Cloud's Vertex AI. The thinking mode, which Anthropic calls **extended thinking**, is reserved for paying users; free accounts get the standard mode.

The price for developers stays where Claude 3.5 Sonnet was: $3 for every million tokens you send and $15 for every million it writes. A token is a chunk of text, often part of a word. The thinking Claude does counts as written tokens, so longer thinking costs more.

Anthropic also says the model turns down harmless requests less often, with 45% fewer unnecessary refusals than its predecessor, and it extended its GitHub integration, which connects your code repositories to Claude, to all plans.

## How hybrid reasoning works

Think of a teacher grading homework. For "What is 7 times 8?" they answer on the spot. For a tricky word problem, they grab scrap paper first. Claude 3.7 Sonnet works the same way, except you tell it which kind of question it is facing.

- **Standard mode** behaves like an upgraded Claude 3.5 Sonnet: quick, conversational answers.
- **Extended thinking** makes Claude write out its reasoning first, checking and correcting itself before it commits to a reply. Anthropic says this helps most with math, physics, coding and following detailed instructions.

The reasoning is visible. In the Claude app you can open a scratch pad and read how Claude got to its answer. Dianne Penn, Anthropic's product and research lead, told TechCrunch that most of it will be shown, though some parts may be hidden for safety reasons.

Developers get a dial. Through the API they can set a **thinking budget**: a cap on how many tokens Claude may spend thinking, anywhere up to its output limit of 128,000 tokens. A small budget is fast and cheap; a large one buys more careful work.

Anthropic also says it tuned the model less for math competitions and more for the everyday tasks businesses give it. On SWE-bench Verified, a test built from real bugs in open-source projects, TechCrunch reports a score of 62.3%, against 49.3% for OpenAI's o3-mini. With extra computing and a more elaborate setup, Anthropic reports 70.3%. On TAU-bench, which simulates an AI agent helping customers of a store, Claude scored 81.2%, ahead of OpenAI's o1 at 73.5%.

## What Claude Code does

Claude Code is a **coding agent**: a program that does not just suggest code but takes actions. It runs in the **terminal**, the text window where programmers type commands. From there, Claude can search and read a project's files, edit them, write and run tests, save the changes to GitHub and use other command-line tools. It describes what it is doing as it goes, and you approve its steps.

In a demo, Anthropic staff simply typed a request in plain English asking Claude to explain how a project was organized. Anthropic says that in its own early testing, Claude Code finished in one go tasks that would normally take a programmer more than 45 minutes by hand.

It is a limited research preview. Access goes to a limited number of people on a first-come, first-served basis, and Anthropic says it will improve the tool over the coming weeks based on how developers use it.

Here is a typical first request a developer might type:

```prompt
Look through this project and explain, in plain language, what each folder does.
Then find the function that sends the weekly email report, write a test for it,
and run the test. Show me any changes before you make them.
```

## Why it matters this week

Reasoning models have been the big story of recent months. OpenAI's o1 and o3-mini, DeepSeek's R1 and others all spend extra time and computing before answering. Claude 3.7 Sonnet brings that ability to Anthropic's lineup, and it tries to remove a common headache: choosing among several models with different costs and strengths. Anthropic has said it would eventually like Claude to decide for itself how long to think.

Claude Code points in a second direction. Instead of copying code from a chat window into your project, you hand Claude the project and watch it work. That is powerful and calls for care: anything an agent changes, you still need to review.

> **Try it:** If you have a paid Claude plan, ask the same tricky question twice, once in standard mode and once with extended thinking. A good test is a logic puzzle or a budget calculation with several steps. Open the thinking, follow the reasoning, and note whether the final answers differ.

## How do you choose between standard and extended thinking in Claude 3.7 Sonnet?

Use standard mode for most everyday work: drafting an email, summarizing a page, brainstorming names, answering a factual question. These tasks rarely improve with extra thought, and a fast answer is more pleasant. Switch to extended thinking when the task has several steps that depend on each other, such as a math problem, a schedule with constraints, a tricky bug, or instructions with many conditions.

A simple rule: if you would need scrap paper to solve it yourself, give Claude the scrap paper too. Anthropic says prompting works much the same in both modes, so you do not need to rewrite your request. Ask clearly, give the details, and then read the visible reasoning. It is often the quickest way to spot where a wrong assumption crept in, whether it is Claude's or a gap in your own question.

For developers, the thinking budget adds a middle ground. Start small, measure quality on your own examples, and raise the budget only where the better answers are worth the extra tokens.

## Sources

- Claude 3.7 Sonnet and Claude Code: https://www.anthropic.com/news/claude-3-7-sonnet
- Anthropic launches a new AI model that 'thinks' as long as you want: https://techcrunch.com/2025/02/24/anthropic-launches-a-new-ai-model-that-thinks-as-long-as-you-want/