Anthropic has fired the latest shot in the AI coding race. The company released Claude Sonnet 4.5 on Monday, billing the artificial intelligence model as the best in the world for coding, real-world computer use and autonomous agents, TechCrunch reported.
The headline number is 77.2 per cent on SWE-bench Verified, the closely watched software-engineering benchmark. Anthropic says the model can run agents continuously for 30 hours, writing code and using tools across multi-step tasks — the kind of endurance that moves AI assistants from autocomplete toward delegated work.
Sonnet 4.5 is priced at $3 per million input tokens and $15 per million output tokens, and arrives with a new native VS Code extension, an Agent SDK, upgraded checkpoints in Claude Code and a research preview called 'Imagine with Claude.'
Anthropic says cyber and biology safeguards were strengthened for the release, an acknowledgement that longer-running agents raise the stakes of misuse as much as of productivity. The model's real test starts now: whether the world's developers, burned before by benchmark claims, find the 30-hour agent as reliable in their repositories as it is in Anthropic's.
