Anthropic Claude Sonnet 4.5 release: state‑of‑the‑art coding model and new developer tools

TL;DR

Claude Sonnet 4.5 is Anthropic’s newest frontier model, delivering the strongest coding, reasoning, and computer‑use abilities to date, and it ships with a suite of product upgrades—including checkpoints in Claude Code, a VS Code extension, new context‑editing tools, and the open‑source Claude Agent SDK.

---

What Claude Sonnet 4.5 is

• Best‑in‑class coding model – leads the SWE‑bench Verified benchmark and can stay focused on complex, multi‑step tasks for over 30 hours.
• Top computer‑use performer – scores 61.4 % on OSWorld, a jump from 42.2 % for Sonnet 4 only four months earlier.
• Improved reasoning and math – shows large gains across public evaluation suites (see benchmark table in the announcement).
• Most aligned frontier model – safety training reduces sycophancy, deception, power‑seeking, and prompt‑injection vulnerabilities.
• Pricing unchanged – $3 / $15 per million input/output tokens, identical to Claude Sonnet 4.

Core Technical Advances

Coding and Long‑Horizon Tasks

• Maintains coherence across 30 + hours of autonomous coding, enabling engineers to hand off months‑long architectural work.
• Parallel tool execution maximises actions per context window, allowing multiple bash commands to run simultaneously.
• Internal benchmarks report a drop from 9 % to 0 % error on Anthropic’s code‑editing test set.

Computer Use (OSWorld)

• Achieves 61.4 % success on real‑world computer tasks such as web navigation, spreadsheet manipulation, and file creation.
• Demonstrated in the Claude for Chrome extension, where the model fills spreadsheets, navigates sites, and generates slides directly in the browser.

Reasoning & Math

• Outperforms prior Claude models on a broad suite of reasoning and mathematics evaluations (see the posted benchmark table).
• Domain‑specific experts in finance, law, medicine, and STEM report dramatically better knowledge and reasoning compared with older models, including Opus 4.1.

Alignment & Safety

• Automated behavioural auditor scores show lower incidence of deceptive, sycophantic, power‑seeking, and delusional outputs.
• Prompt‑injection resistance has been substantially improved for agentic and computer‑use capabilities.
• Deployed under Anthropic’s AI Safety Level 3 (ASL‑3) framework with CBRN classifiers that have been refined to reduce false positives by a factor of ten since their initial release.

Product Enhancements Launched with Sonnet 4.5

Claude Code

• Checkpoints: Users can save progress and instantly roll back to prior states.
• Refreshed terminal UI and a native VS Code extension for seamless development.
• Context editing & memory tool (via the Claude API) enables agents to run longer and handle greater complexity.

Claude Apps

• Integrated code execution and file creation (spreadsheets, slides, documents) directly within conversations.
• Claude for Chrome extension now available to Max users who joined the waitlist.

Claude Agent SDK

• Open‑source infrastructure that powers Claude Code, now available for developers to build custom agents.
• Handles long‑running memory management, permission systems, and sub‑agent coordination.
• Promises similar capability across non‑coding tasks, extending the model’s utility beyond software development.

Early Customer Feedback (Selected Quotes)

"We're seeing state‑of‑the‑art coding performance from Claude Sonnet 4.5, with significant improvements on longer horizon tasks. It reinforces why many developers using Cursor choose Claude for solving their most complex problems."

"Claude Sonnet 4.5 amplifies GitHub Copilot's core strengths – multi‑step reasoning and code comprehension are markedly better, enabling agentic experiences that handle complex, codebase‑spanning tasks."

"Claude Sonnet 4.5 reduced average vulnerability intake time for our security agents by 44 % while improving accuracy by 25 %, helping us lower risk for businesses."

"The model's edit capabilities are exceptional – we went from a 9 % error rate on Sonnet 4 to 0 % on our internal code‑editing benchmark."

"Claude Sonnet 4.5 increased planning performance by 18 % and end‑to‑end eval scores by 12 % for the Devin system – the biggest jump since Claude Sonnet 3.6."

Research Preview: Imagine with Claude

• A temporary demo where Claude generates software live, with no pre‑written code, showcasing real‑time adaptation.
• Available to Max subscribers for five days at claude.ai/imagine.

How to Access Sonnet 4.5

• API: Use model identifier claude-sonnet-4-5 via the Claude API.
• Apps: All paid Claude app plans now include code execution and file creation.
• Developer Platform: Claude Agent SDK and updated platform documentation are publicly available.

Why This Matters

Claude Sonnet 4.5 pushes the frontier of AI‑assisted software development, enabling autonomous agents that can reason, code, and interact with computers at a level previously unseen. The simultaneous release of developer‑focused tools (checkpoints, VS Code extension, Agent SDK) lowers the barrier for building production‑grade AI agents, potentially reshaping how software is built, audited, and maintained across industries.