Anthropic Releases Claude Sonnet 4.5 With Advanced Coding Capabilities

• Launch & Availability (Sept 29, 2025):

  Claude Sonnet 4.5 is available via the Claude API and chatbot at unchanged rates—$3 per million input tokens and $15 per million output tokens.

• Enterprise Adoption:

  Apple and Meta already use Claude models internally. Companies like Cursor, Windsurf, and Replit integrate them into coding apps. Additional “vibe-coding” tools such as Lovable and Devin are cited as ecosystem touchpoints for Claude-based workflows.

• New Tools:

  Launch includes the Claude Agent SDK (TypeScript & Python) for building agents on top of Claude Code, with solutions for memory across long-running tasks, permission systems balancing autonomy with user control, and coordination of sub-agents. Anthropic also published engineering guidance for agent design.

• Expanded Coding Surfaces & Features:

  • Code Interpreter on claude.ai: sandboxed Python and Node.js, with the ability to clone from GitHub and install from NPM/PyPI. Independent testing shows Sonnet 4.5 checking out a repo, running 466 tests in ~168 s, then implementing a tree-structured conversation feature with 22/22 tests passing and packaging deliverables.

  • VS Code extension released; enhanced terminal experience in Claude Code; checkpoints for rollback; code execution & file creation (spreadsheets, slides, docs) directly in chat.

  • Claude for Chrome available to Max users who joined the waitlist; code execution and file creation are available on all paid plans in the apps.

• Benchmark Leadership & Long-Run Tasks:

  Sonnet 4.5 excels on SWE-Bench Verified; external coverage reports a ~77.2% score. In enterprise trials, it maintained autonomous activity for ~30 hours, standing up databases, buying domains, and running SOC 2 checks while building an app. Anthropic also reports 61.4% on OSWorld for real-computer use (up from 42.2% for Sonnet 4 four months earlier) indicating major gains in tool use.

• Reliability, Safety & Governance:

  Anthropic claims reduced sycophancy/deception and stronger defenses against prompt-injection. Sonnet 4.5 ships under ASL-3 protections with CBRN-focused classifiers; Anthropic reports 10× fewer false positives vs the original disclosure (and 2× fewer since Opus 4). If classifiers interrupt a task, conversations can continue with Sonnet 4. The system card includes new behavioral audits and techniques from mechanistic interpretability.

• Pricing Context vs Rivals:

  While Sonnet 4.5 keeps $3 / $15 pricing, Simon Willison notes GPT-5 and GPT-5-Codex at $1.25 / $10, underscoring a capability-vs-price positioning rather than undercutting on cost.

• Wider Availability & Ecosystem:

  Early distribution includes OpenRouter, Cursor, and GitHub Copilot public preview; Anthropic indicates Sonnet 4.5 is “available everywhere today” via API. The Agent SDK and developer platform updates are available to all developers.

• Context & Cadence:

  Released <2 months after Claude Opus 4.1, and less than a month after a $13B funding round, reinforcing Anthropic’s rapid execution amid competition from OpenAI GPT-5 and Google Gemini.

Why This Matters:

Claude Sonnet 4.5 raises the bar for autonomous software development, enabling enterprises to shift from prototyping to fully production-ready AI-built systems. The broader toolchain (Agent SDK, VS Code integration, checkpoints, code execution, Chrome automation, and ASL-3 safety) pushes AI deeper into developer workflows while intensifying competitive pressure on OpenAI and Google across pricing, capability, governance, and agentic reliability.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.