Anthropic Unveils Claude 4.5 Sonnet: “World’s Best Coding Model!”

Throughout the year, Anthropic, the publisher of Claude, has been producing models at a steady pace. They have already released three models: Claude 3.7 Sonnet, Claude 4 Opus & Sonnet, and Claude Opus 4.1. Now, they introduce a fourth version called Sonnet 4.5, which notably includes enhanced performance for programming tasks.

A New Benchmark AI for Coding?

In a modest claim, Anthropic describes Claude Sonnet 4.5 as the “world’s best coding model”. The company highlights its top-tier performance on specialized benchmarks such as SWE-bench Verified, which evaluates a model’s ability to tackle real-world programming problems. Anthropic reports that Claude 4.5 remains focused and coherent for over 30 consecutive hours on complex tasks.

Code is everywhere. It powers every application, spreadsheet, and software tool you use. Being able to utilize these tools and solve complex problems is crucial for success in today’s work environment. Claude Sonnet 4.5 makes all this possible, says Anthropic.

Anthropic also emphasizes two practical additions aimed at enhancing developers’ experience:

• Checkpointing: this feature saves the progress of a project and allows reverting to a previous state. It provides developers with additional security against errors or risky experiments.
• Native Integration in VS Code: this new feature directly connects Claude Code to the most widely used editor among programmers. It simplifies daily work by minimizing switching between tools and enhancing development fluidity.

Agent-Based AI: A More Versatile Tool for Computers

With Claude 4.5, Anthropic further expands the capabilities of its model. The company highlights its advancements on OSWorld, a benchmark that measures an AI’s ability to perform concrete tasks on a computer, such as web navigation, spreadsheet manipulation, or app management. Here, Claude 4.5 achieves a score of 61.4%, an improvement from 42.2% achieved by Sonnet 4 just four months ago.

Several enhancements boost this versatility. Code execution is now possible directly within the Claude application, bringing the tool closer to functioning like a real workstation. The AI can also create files, whether documents, presentations, or spreadsheets. The Chrome extension further allows Claude to interact directly with the browser, enabling it to navigate between different websites or fill in online spreadsheets.

For developers, Anthropic now offers the Claude Agent SDK. Based on the infrastructure of Claude Code, it enables the design of agents capable of handling prolonged tasks, coordinating multiple sub-agents, and finding a balance between autonomy and oversight.

A More Secure and Reliable Model

Anthropic has made significant progress in terms of security. Claude Sonnet 4.5 addresses several concerning behaviors, such as excessive flattery, deceit, power-seeking, or the propensity to encourage delusional thoughts. The model also offers improved resistance to prompt injection attacks that attempt to alter its function.

The publisher has also refined its moderation filters, designed to block sensitive content related to chemical, biological, radiological, or nuclear weapons. These filters sometimes led to overblocking, but their accuracy has now been improved, significantly reducing the number of false positives.

Finally, Claude Sonnet 4.5 is rated at the internal security level ASL-3. To better regulate its behavior, Anthropic uses new evaluation methods like mechanistic interpretability to more accurately anticipate potential model deviations.

Claude Sonnet 4.5 is now available to all users, whether through the API, Claude applications, or Claude Code. It directly replaces Sonnet 4 with no additional cost: the price remains set at $3 per million input tokens and $15 per million output tokens.