Claude Sonnet 4.5 has just been announced and released!
In this article, I’ve compiled the key highlights of Anthropic’s newest model (some of which are genuinely impressive) to give you a clear and concise overview. I’ll save my personal thoughts for the end, especially since I’m currently testing Perplexity Pro and Gemini 2.5 Pro in parallel. Despite Claude being one of my most-used daily tools, this article isn’t sponsored in any way. This blog aims to deliver news as accurately and impartially as possible.
What’s new
“Claude Sonnet 4.5 is the world’s best coding model.” That’s how Anthropic’s blog post opens (https://www.anthropic.com/news/claude-sonnet-4-5), making it crystal clear what this model’s primary focus is: development.
Claude Agent SDK
This is the biggest announcement: an SDK for building agentic applications with Claude is now available. Notably, it’s the same SDK Anthropic’s development team uses for their internal agents. Here’s a brief explainer video: Building agents with the Claude Agent SDK
30 Consecutive hours of coding
Details here are sparse, but Anthropic explains that improvements to memory and context management have enabled Claude to autonomously code for 30 straight hours while maintaining focus on the original objective without losing track.
Context drift as conversations with AI models progress is always frustrating, so I’m eager to test this new capability over the coming days.
Claude Code – Visual Studio Code integration (finally!)
I’d previously written about integrating Claude Code with various IDEs, but those were incomplete implementations, Claude relied on a terminal (or PowerShell) within the IDE to communicate with us.
With Claude 4.5, they’ve finally released a proper Visual Studio Code extension that delivers the visual experience we’re already familiar with from tools like GitHub Copilot, Gemini, and others.
For Visual Studio 2026, we’ll be able to integrate our preferred model directly into Copilot (via API key), allowing us to use it natively within the tool with all the conveniences and integrations Visual Studio 2026 already provides.
Office 365 integration
Microsoft has announced it will integrate new Copilot for Microsoft 365 features powered by Anthropic’s models (including agent mode), rolling out first in Excel and Word, with PowerPoint following in a second phase. While Microsoft bet heavily on OpenAI from the start, these announcements suggest Redmond is broadening its strategy to include alternative models, particularly given recent friction with Sam Altman regarding OpenAI’s direction.
Claude Code – Checkpoint
Imagine you’ve built an application that uses Claude to generate medical reports. You’ve tested everything with Claude Sonnet 4.5 from September 30, 2025, and it works perfectly: the formatting is correct, the tone is appropriate, and user feedback is positive.
Checkpoints will allow you to continue using that specific model version without being affected by Anthropic’s model updates that might alter response styles.
Enhanced security
The new version’s improvements extend to model security, both in terms of response content and from a more technical perspective. On the technical front, they’ve improved detection and mitigation of prompt injection attacks that could force unintended behaviors.
Regarding content, they’ve improved handling of conversations where the model might exhibit excessive compliance, deception, manipulation, or provide dangerous information.
Beyond this, Claude implements AI Safety Level 3 (ASL-3), an intermediate level on their security scale. This level requires specific protocols both before and after model deployment (such as evaluations during training, dangerous content restrictions, and misuse monitoring).
Chrome extension (Max users only)
Users on the Max plan get early access to a plugin that lets them use Claude within Chrome, both for summarizing page content and interacting with it. I expect this feature will roll out to all users with the next model version. Here’s their article for more details: https://www.anthropic.com/news/claude-for-chrome
Imagine with Claude (Max users only)
Also exclusive to Max subscribers, this feature generates websites based on written prompts. The generated sites are purely visual, ideal for exploring layout ideas or page functionality before diving into actual product development. This is powered by agents that create and render the website for you.
Benchmarks
Everything we’ve covered so far gives us a general picture of what’s new, but how does it stack up in the AI arena? How does it compare to its rivals? I won’t include any charts in this article because each model excels in certain areas while underperforming in others. Show here picking graphs would give an incomplete and biased view of a model’s capabilities.
For those interested in deeper analysis, specialized sites like Artificial Analysis are available.
Pricing
The good news is that pricing remains unchanged from the previous model (4.1), and the model is already available to all subscribers.
Conclusions and impressions
As I mentioned at the start, I’ve used Claude as my primary model for about two years, and for development tasks (my main use case), I think it’s an excellent support tool. Recently I’ve been testing both Gemini 2.5 Pro and Perplexity (though the latter focuses more on source-based research). My impression is that Gemini performs significantly better for analysis: when I need to brainstorm or dive deep into complex technical aspects of software or architecture, Gemini delivers more satisfying results. Claude, while also providing quality responses, has a heavy development focus and always includes code in its answers, even when my question is more architectural in nature.
What I’m most curious about is its improved ability to maintain memory and context focused on the objective without degrading conversation quality as the discussion progresses. I’m going to test and compare this with other models!
