Anthropic launches Claude Sonnet 4.5 for agents

Anthropic has launched Claude Sonnet 4.5, an AI model focused on programming, computer use and long-running tasks that require multiple steps. The company says it is its most capable model for building agents, meaning systems that can plan and carry out actions using tools with less supervision.

The model is available today in the Claude apps, Claude Code and the API. For developers, it is identified as claude-sonnet-4-5 and keeps Sonnet 4's pricing: $3 per million input tokens and $15 per million output tokens.

More capability for coding and using tools

Claude Sonnet 4.5 scores 77.2% on SWE-bench Verified, a test that measures whether an AI can solve real problems in software repositories. Anthropic clarifies that this figure corresponds to a specific configuration, with 200,000 reasoning tokens and a limited set of tools.

With a higher-compute configuration and several attempts running in parallel, the company reports a result of 82%. This is not the performance every user will automatically get, but it shows how far the model can go when it has more resources to analyze and check its answers.

It also improves at tasks that involve operating a computer. On OSWorld, an evaluation of actions in real computer environments, Sonnet 4.5 reaches 61.4%, compared with 42.2% for Claude Sonnet 4 four months earlier.

In practice, this allows Claude to browse web pages, fill out spreadsheets, run commands, review files and complete multi-step processes. Anthropic says it has observed sessions lasting more than 30 hours on complex programming tasks, although that figure describes the company's observations, not a guarantee for every use case.

Changes to Claude Code and the API

The update is not limited to the model. Claude Code, Anthropic's product for working with code from a terminal, adds several new features:

• Checkpoints for saving progress and quickly returning to an earlier state.
• A redesigned terminal interface.
• A native extension for Visual Studio Code.
• Improvements for working longer on complex projects.

Anthropic is also introducing the Claude Agent SDK, the set of tools it uses internally to build Claude Code. Developers can use it to create their own agents, including outside programming, with features for managing memory, permissions and multiple subagents that collaborate on the same task.

The API is also getting context editing and a memory tool. Context is the information the model can keep in mind during a session. These features are designed to help agents work on longer tasks without losing track of earlier steps.

More features in the apps

Claude's apps can now run code and directly create files such as spreadsheets, presentations and documents within a conversation. This changes how you use the chatbot: instead of merely explaining how to prepare a report, it can generate the file and modify it based on new instructions.

The Claude for Chrome extension is also opening up to Max users who were on the waitlist. Anthropic presents it as a way to bring computer-use capabilities to the browser, although this type of access requires carefully reviewing which pages the agent visits and what actions it can take.

Safety and limits

Anthropic says Sonnet 4.5 is its most aligned frontier model to date. According to the company, it has reduced behaviors such as excessive flattery, deception, power-seeking and the tendency to reinforce delusional ideas.

The model is being released under AI Safety Level 3 protections, which include filters to detect inputs and outputs potentially related to chemical, biological, radiological or nuclear weapons. These filters can mistakenly block legitimate content. Anthropic says it has reduced these false positives tenfold since its first measurement and twofold since the launch of Claude Opus 4.

The company also says it has improved its defenses against prompt injection attacks, a risk in which a web page or document tries to trick the agent into ignoring the user's instructions. Even so, giving an AI access to a browser, files or a terminal still requires limited permissions and supervision.

As a temporary demonstration, Anthropic is offering Imagine with Claude to its Max subscribers for five days. The system generates software in real time based on the user's instructions, without starting from a fully prepared application.

For you, the main takeaway is simple: Claude Sonnet 4.5 is trying to move from answering questions to completing entire jobs, especially in programming, analysis and file management. What matters will be seeing how consistently that autonomy works outside demonstrations and how the controls evolve as these agents gain access to more tools and data.