September 30, 2025 – Anthropic has officially launched Claude Sonnet 4.5, the latest iteration of its advanced AI programming model. This model can maintain focus on complex, multi-step tasks for over 30 continuous hours, topping benchmarks in programming, computational reasoning, and AI agent performance. It surpasses even GPT-5 in reasoning, mathematics, and agent-driven coding tasks.
Claude Sonnet 4.5 Pricing
Claude Sonnet 4.5 maintains the same pricing model as its predecessor:
-
Input: $3 per million tokens
-
Output: $15 per million tokens
New Features: Claude Code & Agent SDK Upgrades
Claude Code now includes:
-
Checkpointing: Save progress with instant rollback capabilities.
-
Updated Terminal Interface: Enhanced for usability and efficiency.
-
Native VS Code Extension: Allows developers to embed Claude Code directly into their IDE.
Anthropic also released the Claude Agent SDK, enabling developers to leverage the core architecture behind Claude Code for custom applications. Meanwhile, Claude API adds context editing and memory tools, allowing AI agents to handle more complex workflows. Tasks like code execution and file generation (spreadsheets, slides, documents) are now seamlessly integrated into conversational workflows.
All of these updates are available today for public beta on the Claude Developer Platform, Amazon Bedrock, and Google Cloud Vertex AI.
1. Sustained Focus for 30+ Hours: Claude Sonnet 4.5 Outperforms GPT-5
In the SWE-bench Verified test for real-world programming capability, Claude Sonnet 4.5 ranks #1, showing it can handle complex multi-step tasks continuously for more than 30 hours.
On the OSWorld benchmark, which evaluates actual computer operation skills, Claude Sonnet 4.5 scored 61.4%, a significant jump from Sonnet 4’s 42.2% just four months ago. The model can fully automate browser-based tasks including website navigation, form-filling, and end-to-end workflow execution.
Experts in law, finance, medicine, and STEM fields confirm that Claude Sonnet 4.5 demonstrates marked improvements in professional knowledge and reasoning compared to previous models, including Opus 4.1.
Anthropic emphasizes that Claude Sonnet 4.5 is not only their most capable model but also their most ethically aligned AI system. Advanced safety training reduces undesirable behaviors such as flattery, deception, power-seeking, or promoting delusional thinking.
2. Native VS Code Extension and Upgraded Claude Code Features
Claude Code introduces several upgrades:
-
Native VS Code Extension: Developers can now use Claude Code directly inside their IDE, with a sidebar panel and inline diff tracking for real-time code changes.
-
Terminal Interface 2.0: Improved visualization and searchable command history enhance user experience.
-
Checkpoint System: Automatically saves code before each change, allowing instant rollback with a double-tap of Esc or
/rewindcommand. Users can restore code, conversation history, or both.
For teams building custom AI workflows, the Claude Agent SDK provides core tools, context management, and permission frameworks. It now supports sub-agents and hooks for more flexible workflow adaptation.
3. Performance Boost and Advanced Context Management
Claude Developer Platform introduces Context Editing and Memory Tools:
-
Context Editing: Automatically removes outdated tool calls and results when token limits are reached, keeping conversations efficient and extending autonomous task execution.
-
Memory Tools: Store and retrieve information outside the active context window, building persistent knowledge across sessions.
With these tools, Claude Sonnet 4.5 can handle full codebases, analyze hundreds of documents, and maintain extensive tool interaction history. Internal tests show performance improvements up to 39% and token usage reduced by 84% when using context management effectively.
Conclusion: A More Complete AI Development Ecosystem
Claude Sonnet 4.5 represents a major leap from model to toolchain. With over 30 hours of sustained focus, enhanced Claude Code, Agent SDK, and advanced context management, Anthropic provides a robust ecosystem for building intelligent agents, capable of tackling complex, real-world workflows with unprecedented efficiency.
