Claude Sonnet 4.5 Is Anthropic's Safest Ai Model Yet

In May, Anthropic announced 2 caller AI systems, Opus 4 and Sonnet 4. Now, little than six months later, nan institution is introducing Sonnet 4.5, and calling it nan champion coding exemplary successful nan world to date. Anthropic's ground for that declare is simply a action of benchmarks wherever nan caller AI outperforms not only its predecessor but besides nan much costly Opus 4.1 and competing systems, including Google's Gemini 2.5 Pro and GPT-5 from OpenAI. For instance, successful OSWorld, a suite that tests AI models connected real-world machine tasks, Sonnet 4.5 group a grounds people of 61.4 percent, putting it 17 percent points supra Opus 4.1. 

At nan aforesaid time, nan caller exemplary is tin of autonomously moving connected multi-step projects for much than 30 hours, a important betterment from nan 7 aliases truthful hours Opus 4 could support astatine launch. That's an important milestone for nan type of agentic systems Anthropic wants to build. 

Perhaps much importantly, nan institution claims Sonnet 4.5 is its safest AI strategy to date, pinch nan exemplary having undergone "extensive" information training. That training translates to a chatbot Anthropic says is "substantially" little prone to "sycophancy, deception, power-seeking and nan inclination to promote illusion thinking" — each imaginable exemplary traits that person landed OpenAI successful basking h2o successful caller months. At nan aforesaid time, Anthropic has strengthened Sonnet 4.5's protections against punctual injection attacks. Due to nan sophistication of nan caller model, Anthropic is releasing Sonnet 4.5 nether its AI Safety Level 3 framework, meaning it comes pinch filters designed to forestall perchance vulnerable outputs related to prompts astir chemical, biologic and atomic weapons.  

With today's announcement, Anthropic is besides rolling retired value of life improvements crossed nan Claude merchandise stack. To start, Claude Code, nan company's celebrated coding agent, has a refreshed terminal interface, pinch a caller characteristic called checkpoints included. As you tin astir apt conjecture from nan name, they let you to prevention your advancement and rotation backmost to a erstwhile authorities if Claude writes immoderate funky codification that isn't rather moving for illustration you imagined it would. File creation, which Anthropic began rolling retired astatine nan commencement of nan month, is now disposable straight successful conversations pinch nan chatbot, and if you joined nan waitlist Claude for Chrome, you tin commencement utilizing nan hold today.   

API pricing for Sonnet 4.5 remains astatine $3 per 1 cardinal input tokens and $15 for nan aforesaid magnitude of output tokens. The merchandise of Sonnet 4.5 caps disconnected a beardown September for Anthropic. Just 1 time aft Microsoft added Claude models to Copilot 365 last week, OpenAI admitted its rival offers nan champion AI for work-related tasks.