Just now, Claude Sonnet 4.5 was grandly released, and a new king of programming has arrived.

Early this morning Beijing time, Anthropic officially released Claude Sonnet 4.5.

This version is defined by Claude as the world's most powerful code model, and it also shows significant breakthroughs in agent construction, computer usage, reasoning, and mathematical abilities.

Code is everywhere. It drives applications, spreadsheets, and various tools.

The core of modern work lies in using these tools and solving complex problems, and Sonnet 4.5 makes all of this more efficient and reliable.

New Features and Product Upgrades

Along with Sonnet 4.5, Anthropic has carried out a large-scale update to the entire Claude product line:

Claude Code has added the long-awaited "Checkpoints" feature, which allows users to save progress at any time and roll back to an earlier state with one click. At the same time, the terminal interface has been updated, and a native VS Code plugin has been launched.

Claude API has added context editing functionality and memory tools, enabling agents to run longer and handle more complex tasks.

Claude apps now support directly executing code and generating files (including spreadsheets, slideshows, and documents) in conversations.

Claude for Chrome Extension is now available to Max users on the previous waiting list.

The developer community has also welcomed a new core resource: Claude Agent SDK.

This toolkit opens up the underlying infrastructure that drives Claude Code, providing the basic capabilities for everyone to build agents.

Cutting-Edge Performance and Benchmark Results

Sonnet 4.5 has performed excellently in multiple authoritative tests:

SWE-bench Verified: It has reached the latest optimal level in real software coding ability tests and can maintain focused execution for more than 30 hours in complex multi-step tasks.

OSWorld: In the benchmark simulating real computer tasks, it ranked first with a score of 61.4%, while Sonnet 4 was still at 42.2% four months ago.

The video demonstrates the ability of Claude for Chrome to simulate computer tasks.

Reasoning and Mathematics: It has significantly led in a series of public evaluations, especially in knowledge and reasoning in professional fields such as finance, law, medicine, and STEM, far surpassing the previous Opus 4.1.

These improvements are not only reflected in the data but also highly praised by professional users who had early access:

Claude Code can program autonomously for more than 30 hours, helping engineers complete architecture work that originally took months in a very short time. ——Sean Ward, CEO of iGent AI

In the internal code editing test of replit, the error rate dropped directly from 9% of Sonnet 4 to 0%. ——Michele Catasta, President of replit

Stronger Alignment, Safer

Sonnet 4.5 is the Claude model with the best alignment to date.

Through improved capabilities and extensive security training, the model has achieved obvious results in reducing behaviors such as pandering, deception, power - seeking, and false encouragement.

In the scenarios of agent and computer usage, the team has also specifically strengthened the defense against prompt injection attacks.

These improvements have been publicly presented in the form of a detailed evaluation in the System Card, in which interpretability technology is introduced for the first time to verify security.

Sonnet 4.5 is released according to the AI Safety Level 3 (ASL - 3) framework, which includes a series of protective measures, such as a classifier filtering mechanism for chemical, biological, radiological, and nuclear - related content.

Although these filters may occasionally misjudge normal content, compared with Sonnet 4, the false positive rate has been reduced by ten times and will continue to be optimized in the future.

Claude Agent SDK

In the past six months, the Claude team has continuously iterated on Claude Code and explored the core challenges in agent design, including memory management for long - term tasks, the balance of permission systems, and the collaboration methods of sub - agents.

Now, these capabilities are fully open to developers through the Claude Agent SDK.

The SDK is not only suitable for coding but also supports a wider range of application scenarios.

It provides the same infrastructure as Anthropic's internal tools for building autonomous agents for complex tasks.

Research Preview: "Imagine with Claude"

As an additional experiment, Anthropic has released a temporary research preview - "Imagine with Claude".

Imagine With Claude generates software demonstration videos in real - time.

In this experience, Claude will generate software in real - time without preset functions or pre - written code. Users can directly see the process of Claude's instant creation and adaptation during the interaction.

This feature is available to Max subscribers for 5 days and can be experienced at claude.ai/imagine.

New API Features

Sonnet 4.5 also brings important updates at the API level:

Memory Tools (Beta): Support storing and retrieving information outside the context window to help maintain knowledge and state across sessions.

Context Editing: Automatically clean up old tool calls and results to avoid context overflow.

New Stop Reason Prompt: Adds model_context_window_exceeded to more clearly identify generation termination due to context limitations.

Tool Parameter Improvement: Fixes the problem of the trailing line break of string parameters being truncated, ensuring accurate format transmission.

Token Counting Optimization: The system will automatically add optimized tokens, but users will not be charged for these system - injected parts.

Pricing and Usage Channels

The price remains the same as that of Sonnet 4:

Input: $3 per million Tokens

Output: $15 per million Tokens

Sonnet 4.5 can be used through the following channels:

Claude API: claude - sonnet - 4 - 5 - 20250929

Amazon Bedrock: anthropic.claude - sonnet - 4 - 5 - 20250929 - v1:0

Google Cloud Vertex AI: claude - sonnet - 4 - 5@20250929

It is also available on the Claude.ai and Claude Code platforms.

Upgrade Guide

Developers currently using Sonnet 4 only need to update the model name to claude - sonnet - 4 - 5 - 20250929 to complete the migration.

All existing API calls remain valid. It is recommended to enable new features such as memory tools and context cleaning at the same time to fully leverage the model's performance.

It should be noted that Sonnet 4.5 no longer allows users to specify both the temperature and top_p parameters simultaneously. Users should choose one of them.

A New Round of AI Programming Wars is Coming

Claude Sonnet 4.5 has achieved an all - round leap in coding, agent capabilities, computer usage, and security.

It is not only a new powerful tool in developers' toolkits but also a strong platform that can provide professional support in fields such as finance, medicine, and scientific research.

Through this version, Anthropic continues to promote the balanced development of AI in cutting - edge capabilities and security.

The new king of programming has arrived. A new round of the "hundred - model war" centered around programming is about to begin, and Claude Sonnet 4.5 is about to become the new benchmark for everyone.

Reference Materials:

https://www.anthropic.com/news/claude - sonnet - 4 - 5

https://docs.claude.com/en/docs/about - claude/models/whats - new - sonnet - 4 - 5

https://assets.anthropic.com/m/12f214efcc2f457a/original/Claude - Sonnet - 4 - 5 - System - Card.pdf

This article is from the WeChat official account "New Intelligence Yuan". The author is New Intelligence Yuan, and the editor is Allen. It is published by 36Kr with authorization.