- It performs 61,4% in OSWorld and leads in SWE-bench Verified
- Handles complex tasks for more than 30 hours and generates up to 64.000 tokens
- Updates to Claude Code and the new Claude Agent SDK for agents
- Enhanced security (ASL-3) and same price: $3/$15 per million tokens
Anthropic has released Claude Sonnet 4.5, an evolution focused on programming, agents, and computer control that seeks to consolidate the platform in professional environments. In a landscape with high-level rivals, the company describes this release as its more refined and useful model for engineering tasks to date.
The new version builds on the track record of the Sonnet family, which had already improved reasoning and coding in previous iterations. Building on that foundation, 4.5 aims to expand the practical scope with advancements in persistence of attention, tool use, and productivity, maintaining a prudent strategy in security and alignment.
Key capabilities and performance improvements
According to Anthropic, Claude Sonnet 4.5 is capable of maintaining focus for more than 30 hours on complex tasks. and multi-step, which favors long projects where continuity of context is required. It also supports outputs of up to 64.000 tokens in a single response, and offers controls to adjust the “thinking time” before responding, balancing speed and detail as needed.
In real tasks in front of the computer, The company reports a 61,4% in OSWorld, a notable jump from its predecessor's 42,2% in this same test.In practical scenarios, the model can browse the web, complete spreadsheets, and perform actions in desktop applications from the Chrome extension, reducing continuous user monitoring.
The land of Programming concentrates most of the improvements. In the SWE-bench Verified evaluation, which focused on coding applied to real-world projects, Sonnet 4.5 leads the way with 77,2% (with configurations that increase the number under parallel computing). Anthropic proposes that the model cover the entire development cycle: planning, implementation, refactoring, and maintenance of large code bases.
Beyond pure development, Anthropic identifies uses that require prolonged flows and coordination of steps.From cybersecurity and finance to office productivity and research using internal and external data. In these contexts, the promise lies in more stable agents capable of sustaining long-term work without losing consistency.
Developer Tools and Ecosystem
The launch comes accompanied by What's new at Claude Code: checkpoints to save progress and return to previous states, such as version historya whirlpool bath, revamped terminal interface, native extension for Visual Studio Code and improvements to context and memory editing via the API to run longer tasks.
Anthropic also premieres the Claude Agent SDK, which replicates the infrastructure the company uses to build its own agentsThe kit offers tools for long-term memory, permission systems, and subagent coordination, facilitating the creation of automated solutions that cooperate toward common goals and secure connectivity with tools such as wire guard.
As a complement, The firm temporarily enables “Imagine with Claude”, a demonstration that allows us to observe how the model generates software in real time No predefined code. This preview, available for a limited time to Max users, illustrates the model's potential for interactive creation.
Security, alignment and resilience
Anthropic includes Sonnet 4.5 in its protection level AI Safety Level 3 (ASL-3), with filters trained to detect dangerous content, especially those related to CBRN risks. The company claims to have reduced false positives by a factor of ten compared to the initial version of these classifiers, and offers Continuity of conversation with Sonnet 4 if a security lockout occurs.
In parallel, the company ensures that The model reduces unwanted behaviors such as flattery or deceptive responses and strengthens defenses against attempts to prompt injectionThese measures point to a use more reliable in corporate environments, where the execution of automated actions requires controls and traceability.
Availability, platforms and prices
Claude Sonnet 4.5 is available at Claude.ai (web, iOS and Android) and for developers via the Claude Developer Platform, with integration into services such as Amazon Bedrock and Google Cloud Vertex AI. The free plan operates with a session limit that resets every five hours and a variable number of messages on demand. Prices remain the same.: $3 per million input tokens and $15 per million output tokens.
Among the new access features, Claude's Chrome extension is rolling out to Max users. previously registered on the waiting list. Although the benchmarks suggest substantial improvements compared to previous iterations, Anthropic notes that actual performance depends on the use case and the reasoning budget configured for each task.
With a combination of advances in coding, greater autonomy for agents, and a stricter focus on security, Claude Sonnet 4.5 is positioned as a solid option for technical teams that need continuity and control in long processes, maintaining stable costs and compatibility with Anthropic's already deployed ecosystem.
I am a technology enthusiast who has turned his "geek" interests into a profession. I have spent more than 10 years of my life using cutting-edge technology and tinkering with all kinds of programs out of pure curiosity. Now I have specialized in computer technology and video games. This is because for more than 5 years I have been writing for various websites on technology and video games, creating articles that seek to give you the information you need in a language that is understandable to everyone.
If you have any questions, my knowledge ranges from everything related to the Windows operating system as well as Android for mobile phones. And my commitment is to you, I am always willing to spend a few minutes and help you resolve any questions you may have in this internet world.