Anthropic Launches Claude Sonnet 4.5, Touts It as ‘Best Coding Model in the World’

Anthropic on Monday announced the release of Claude Sonnet 4.5, its latest AI model for coding and agent-based tasks. The company said the model demonstrates improvements in reasoning, math, and long-duration task management.

“Claude Sonnet 4.5 is the best coding model in the world. It’s the strongest model for building complex agents,” the company said in its blog post. “It’s also the best model at using computers and shows substantial gains in reasoning and math.”

The model is available via the Claude API at the same pricing as Sonnet 4, $3 per million tokens for standard use and $15 per million for extended use.

Anthropic said that the Claude API has added context editing and memory tools to support longer tasks, and the Claude apps now allow code execution and file creation directly within conversations. Anthropic also released the Claude for Chrome extension for Max users on the waitlist.

Claude Sonnet 4.5 is also integrated into Claude Code, which now includes checkpoints to save progress and roll back to previous states, a refreshed terminal interface, and a native VS Code extension. 

Developers can access the Claude Agent SDK, which provides the infrastructure used internally to build Claude Code. “The Agent SDK gives you the same foundation to build something just as capable for whatever problem you’re solving,” the spokesperson said.

Claude Sonnet 4.5 achieved top scores on the SWE-bench Verified evaluation, which tests real-world software coding skills. On OSWorld, a benchmark for real-world computer tasks, the model scored 61.4%, up from 42.2% for Claude Sonnet 4. Early users reported improved performance across finance, law, medicine, and STEM domains.

The company emphasised safety and alignment improvements, noting reductions in misaligned behaviour such as sycophancy, deception, and power-seeking. The model is released under Anthropic’s AI Safety Level 3 framework, which includes classifiers to flag potentially dangerous content.

Anthropic also introduced a temporary research preview, “Imagine with Claude,” which allows users to see the model generate software in real time. It is available to Max subscribers for five days at claude.ai/imagine.