New Claude Sonnet 4.5 Just Broke EVERY Limit We Knew (So Powerful It’s Scary)

Anthropic's new Claude Sonnet 4.5 exhibits unprecedented endurance and capability, achieving state-of-the-art coding performance and operating for 30 hours straight on complex projects.

• Claude Sonnet 4.5 performed an unprecedented 30-hour coding run, demonstrating superior endurance and state-of-the-art coding benchmarks.

• The model is integrated into major platforms like GitHub Copilot and Microsoft Office 365, with a new Agent SDK for custom AI development.

• Sonnet 4.5 features enhanced safety under the AI Safety Level 3 framework, with advanced content filters and reduced undesirable behaviors.

Claude Sonnet 4.5 marks a significant leap in AI capabilities, demonstrating a 30-hour autonomous coding run and delivering state-of-the-art results on coding benchmarks like SWEBenchVerified. This model shows substantial improvements in efficiency across various enterprise workflows, from code planning and financial insights to vulnerability triage, and is being widely integrated into major developer and productivity platforms. Anthropic has also released a new Agent SDK, enabling developers to build sophisticated AI agents with managed virtual machines and enhanced memory systems.

Unprecedented Endurance & Performance

• 00:00:02 Anthropic's Claude Sonnet 4.5 has achieved a remarkable 30-hour continuous coding run without losing focus, a significant improvement over previous versions that lost steam after about seven hours. This AI now holds a state-of-the-art ranking on SWEBenchVerified for coding and improved its OSWorld score to 61.4% from 42% in four months. Sonnet 4.5 also demonstrated an 18% boost in code planning and a 12% improvement in end-to-end results for AI development, along with investment-grade insights in finance and a 44% reduction in vulnerability triage time with improved accuracy in security.

Expanded Integrations & Product Updates

• 00:01:46 Sonnet 4.5 comes with a host of product updates and wider integrations, including checkpoints in Claude Code for instant rollbacks and a redesigned terminal for smoother workflows. A native VS Code extension has been released, allowing Sonnet to integrate seamlessly into developers' preferred environments, while the Claude API now features a memory system and context editing for longer, more complex sessions. The model is actively deployed in GitHub Copilot for pro, enterprise, and business customers, and is being integrated into Microsoft Office 365 Copilot with new agent modes for Excel and Word, extending its reach beyond Anthropic's own applications.

New Agent SDK & Memory Management

• 00:02:35 A key highlight for developers is the new Claude Agent SDK, which provides access to Anthropic's internal infrastructure for Claude Code, including managed virtual machines, memory modules, and APIs for context and editing. This allows users to build their own agents capable of running scripts for hours, remembering history across sessions, and coordinating sub-agents. Anthropic spent over six months refining memory management for long tasks and designing frameworks that balance autonomy with user oversight, all now packaged within the SDK.

Enhanced Safety & Alignment

• 00:05:28 Claude Sonnet 4.5 is described as Anthropic's most aligned Frontier model, shipping under their AI Safety Level 3 framework which combines high capability with strict safeguards. This includes advanced filters for chemical, biological, radiological, and nuclear content, alongside reinforced defenses against prompt injection attacks. Internal audits show significant reductions in undesirable behaviors like deception and power-seeking, with false positive content flags cut by a factor of 10 compared to Opus 4. Anthropic also utilized mechanistic interpretability tools for the first time to analyze the model's internal reasoning, indicating a deeper commitment to alignment research.