Claude Sonnet 4.5: Anthropic’s new LLM can code for 30 hours straight

They’re wasting no time: Less than a month after the major funding round of $13 billion, AI company Anthropic has today released Claude Sonnet 4.5 – a new language model that, according to the company, is designed to be particularly leading in the field of software development. The model thus positions itself in direct competition with OpenAI’s GPT-5 and Google’s Gemini 2.5, aiming to beat them especially in programming capabilities.

Anthropic has already established a very strong reputation with Claude regarding coding and is integrated into vibe-coding tools like Lovable, Cursor, or Devin. With Claude Sonnet 4.5, the company now wants to secure this position against OpenAI and Google (more on this here).

Performance Data and Benchmarks

According to Anthropic, Claude Sonnet 4.5 achieves top scores among available AI models on the SWE-bench Verified test, which measures practical programming capabilities. In controlling computer systems, the model achieves a score of 61.4 percent on the OSWorld benchmark – a significant increase compared to its predecessor Claude Sonnet 4, which stood at 42.2 percent four months ago.

The company states that the model can work with focus on complex multi-step tasks over more than 30 hours. According to Anthropic, improvements are also evident in logical reasoning and mathematics, as well as in domain-specific knowledge in areas such as finance, law, medicine, and natural sciences.

Product Integration and New Features

Parallel to the model’s release, Anthropic has made several product updates. In Claude Code, the company’s development environment, checkpoints have been introduced that allow users to return to earlier work states. Additionally, a native extension for Visual Studio Code has been provided.

In the Claude applications, users can now execute code directly in the conversation and create files such as spreadsheets, presentations, and documents. The Chrome extension has been made available to Max subscribers.

With the Claude Agent SDK, Anthropic provides developers with the infrastructure that is also used for its own products. The SDK is designed to enable the construction of complex AI agents and includes solutions for memory management in long-running tasks, permission systems, and the coordination of multiple sub-agents.

Safety and Alignment

Anthropic describes Claude Sonnet 4.5 as the company’s most strongly aligned model to date. According to the company, problematic behaviors such as excessive agreement, deception, or encouraging unrealistic expectations have been reduced. Particular progress has been made in defending against prompt injection attacks, a security vulnerability in AI systems.

The model operates under Anthropic’s ASL-3 safety protocols, which include filters for detecting potentially dangerous content, particularly in connection with chemical, biological, radiological, and nuclear weapons. Anthropic states that it has reduced false detections by a factor of ten since the initial description.

Availability and Pricing

Claude Sonnet 4.5 is available immediately via the Claude API. The pricing remains identical to the predecessor model Claude Sonnet 4 at $3 for one million input tokens and $15 for one million output tokens. Additionally, Anthropic is offering a research preview called “Imagine with Claude” for Max subscribers for the next five days, which is designed to demonstrate real-time software generation.