Claude Sonnet 4.5 is Anthropic’s safest AI model yet

In Could, Anthropic introduced two new AI techniques, Opus 4 and Sonnet 4. Now, lower than six months later, the corporate is introducing Sonnet 4.5, and calling it the perfect coding mannequin on this planet to this point. Anthropic’s foundation for that declare is a collection of benchmarks the place the brand new AI outperforms not solely its predecessor but additionally the dearer Opus 4.1 and competing techniques, together with Google’s Gemini 2.5 Pro and GPT-5 from OpenAI. For example, in OSWorld, a collection that exams AI fashions on real-world pc duties, Sonnet 4.5 set a report rating of 61.4 p.c, placing it 17 proportion factors above Opus 4.1. 

On the similar time, the brand new mannequin is able to autonomously engaged on multi-step tasks for greater than 30 hours, a big enchancment from the seven or so hours Opus 4 may preserve at launch. That is an necessary milestone for the kind of agentic techniques Anthropic desires to construct. 

Maybe extra importantly, the corporate claims Sonnet 4.5 is its most secure AI system to this point, with the mannequin having undergone “intensive” security coaching. That coaching interprets to a chatbot Anthropic says is “considerably” much less liable to “sycophancy, deception, power-seeking and the tendency to encourage delusional pondering” — all potential mannequin traits which have landed OpenAI in hot water in recent months. On the similar time, Anthropic has strengthened Sonnet 4.5’s protections towards immediate injection assaults. As a result of sophistication of the brand new mannequin, Anthropic is releasing Sonnet 4.5 beneath its AI Security Stage 3 framework, which means it comes with filters designed to forestall probably harmful outputs associated to prompts round chemical, organic and nuclear weapons.  

With at present’s announcement, Anthropic can be rolling out high quality of life enhancements throughout the Claude product stack. To begin, Claude Code, the corporate’s fashionable coding agent, has a refreshed terminal interface, with a brand new characteristic known as checkpoints included. As you possibly can in all probability guess from the title, they permit you to save your progress and roll again to a earlier state if Claude writes some funky code that is not fairly working such as you imagined it could. File creation, which Anthropic began rolling out at the start of the month, is now obtainable instantly in conversations with the chatbot, and should you joined the waitlist Claude for Chrome, you can begin utilizing the extension at present.   

API pricing for Sonnet 4.5 stays at $3 per a million enter tokens and $15 for a similar quantity of output tokens. The discharge of Sonnet 4.5 caps off a robust September for Anthropic. Simply in the future after Microsoft added Claude models to Copilot 365 final week, OpenAI admitted its rival affords the perfect AI for work-related duties.