# What is Claude Sonnet 4.5, and what is the Claude Agent SDK?

Date: 2025-09-30 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-sonnet-4-5/

**Answer:** Claude Sonnet 4.5 is Anthropic's new AI model, released on September 29, 2025, aimed at coding, computer use and long-running agents; Anthropic reports it can stay on a complex task for more than 30 hours and it keeps Sonnet 4's price. Alongside it, the Claude Agent SDK gives developers the same building blocks Anthropic uses for Claude Code, so they can build their own agents.

Yesterday, Monday, September 29, 2025, Anthropic released **Claude Sonnet 4.5**. Sonnet is Anthropic's middle model family, sitting between the small, fast Haiku and the large Opus. This time the middle model leads: Anthropic calls Sonnet 4.5 its strongest model so far for writing code, building agents and operating a computer, and it recommends it as the upgrade for every use.

It is available to everyone in the Claude apps and to developers under the name `claude-sonnet-4-5`, at the same price as Sonnet 4: $3 per million input tokens and $15 per million output tokens. A **token** is a small chunk of text, roughly part of a word.

## What Anthropic released

The model came with a wave of product changes:

| Where | What is new |
|---|---|
| Claude Code | **Checkpoints** that save progress so you can roll back instantly, a refreshed terminal look, and a native VS Code extension |
| Claude API | **Context editing** and a **memory tool**, so agents can run longer without losing track |
| Claude apps | Code execution and file creation (spreadsheets, slides, documents) inside the chat, on paid plans |
| Claude for Chrome | The browser extension opens to Max subscribers who joined last month's waitlist |
| Claude Agent SDK | The toolkit behind Claude Code, now open to all developers |

There is also a playful research preview, **Imagine with Claude**, in which Claude generates software on the fly as you click, with no code written in advance. It is open to Max subscribers for five days.

## How much better is it?

Anthropic's own numbers center on practical work:

- **SWE-bench Verified**, a test that asks a model to fix real bugs from open-source projects and checks the fixes with tests: 77.2%, averaged over ten runs. With extra computing power, sampling several attempts in parallel and picking the best, it reaches 82.0%.
- **OSWorld**, a test of real tasks on a computer, such as working in apps and websites: 61.4%. Four months ago Sonnet 4 led this test with 42.2%.
- **Stamina:** Anthropic says it has seen the model stay on a single complex, multi-step job for over 30 hours. Axios compares that with about seven hours of independent work for Claude Opus 4.

Anthropic adds that experts in finance, law, medicine and science found its knowledge and reasoning much stronger than older models, including Opus 4.1. In an interview with Axios, Anthropic product lead Scott White said the model is strong at work that calls for research and diligence, and described the arc as Claude moving from assistant to collaborator to a fully autonomous agent.

These figures come from Anthropic. Independent testing will show how well they hold up on everyday tasks.

## What the Claude Agent SDK is

An **SDK**, or software development kit, is a package of ready-made code that developers build on. An **agent** is an AI that pursues a goal over many steps, choosing tools as it goes.

Anthropic has been improving Claude Code for over half a year, and along the way it worked through problems every agent faces: how to manage memory over a long task, how to ask permission without becoming tiresome, and how to coordinate **subagents**, helper agents that each take part of a job. The Claude Agent SDK hands that same infrastructure to developers, and Anthropic stresses that it is useful well beyond coding.

Here is the kind of agent a small team might build on it:

```prompt
You are a support-ticket assistant. For each new ticket:
1. Read the ticket and search our help-center files for the answer.
2. If you find it, draft a reply that quotes the relevant help article.
3. If the ticket mentions a refund or a legal threat, do not reply;
   flag it for a human with a one-line summary.
Keep a log file of every ticket you handle and what you did.
```

The SDK supplies the plumbing: reading files, keeping the log as memory, and the permission checks that stop the agent before sensitive steps.

> **Try it:** Take a task you would normally break into several chats, such as cleaning up a messy spreadsheet and writing a summary of it. Give it to Claude Sonnet 4.5 in one go with numbered steps and ask it to report what it finished after each step.

## Safety and alignment

Anthropic calls Sonnet 4.5 its most **aligned** frontier model, meaning its behavior most closely matches what its makers intend. It reports fewer problem behaviors such as **sycophancy** (telling users what they want to hear), deception and power-seeking, and better defenses against **prompt injection**, hidden instructions in web pages or files that try to hijack an agent. Its system card, the technical report on its testing, includes for the first time checks based on interpretability research, which studies what happens inside the model.

The model is released under Anthropic's **AI Safety Level 3** protections. These include classifiers, automatic filters that watch for dangerous requests about chemical, biological, radiological or nuclear weapons. Anthropic admits the filters sometimes flag ordinary content; when that happens you can continue the conversation with Sonnet 4. It says false alarms are now ten times rarer than when the filters were first described.

## How should you put Claude Sonnet 4.5 to work?

If you already use Claude, switch to Sonnet 4.5: it replaces Sonnet 4 directly at the same price, in the apps, the API and Claude Code. For everyday writing and analysis you will mainly notice sharper answers. The bigger gains appear on long, structured jobs.

Give it work the way you would brief a capable colleague: the goal, the limits, and how to check the result. In Claude Code, try a multi-file change and use checkpoints so you can roll back if a direction goes wrong. In the Claude apps on a paid plan, ask it to produce an actual spreadsheet or slide deck rather than a description of one.

Developers who want their own agent should start small with the Claude Agent SDK: one clear task, a log file, and a rule for when the agent must stop and ask a person. Review its work closely at first, then widen its scope as you learn where it is reliable.

## Sources

- Introducing Claude Sonnet 4.5: https://www.anthropic.com/news/claude-sonnet-4-5
- Anthropic's latest Claude model can work for 30 hours on its own: https://www.axios.com/2025/09/29/anthropic-claude-sonnet-coding-agent