🧭 Claude Sonnet 3.7 & Haiku 3.5 Reach End of Life — Migration Guide

As previously announced in the deprecation schedule, Claude Sonnet 3.7 and Claude Haiku 3.5 have today reached end of life on the direct API. API calls targeting these model IDs will now return a model_deprecated error. AWS Bedrock and Google Vertex AI maintain their own deprecation schedules, with both platforms offering a 30-day grace period for existing deployments — check your provider's documentation for the exact cutover date.

Recommended migration paths

• Sonnet 3.7 → Sonnet 4.6 — the natural upgrade; improves on every benchmark and maintains the same pricing tier. Most prompts work without modification
• Haiku 3.5 → Haiku 4.5 — the current lightweight model; significantly faster and more capable than Haiku 3.5 at the same price point
• Model version pinning tip — Anthropic recommends using dated model IDs (e.g., claude-sonnet-4-6-20260217) in production rather than the generic claude-sonnet-4-6 alias, which will auto-update when Sonnet 4.7 launches

🧭 Claude Code Security Enters Limited Preview & Sets New OSWorld Score

Anthropic has opened a limited preview of Claude Code Security — a specialised mode that wraps Claude Code's agent execution in a hardened sandboxed environment and equips it with security-focused tooling: static analysis integrations, secret scanning, dependency vulnerability lookups, and a structured findings report format. The preview is invite-only and targeted at enterprise security teams and pentest firms interested in using AI agents for code audit workflows.

Alongside the preview announcement, Anthropic has published results showing its computer-use agent achieving a new high score on the OSWorld benchmark — a standardised evaluation of GUI-based computer operation tasks. The agent, built on Opus 4.6 with the latest computer-use tool, scored 72.4% on the full OSWorld task suite, meaningfully ahead of the previous state-of-the-art. Anthropic attributes the gains primarily to improved screenshot comprehension and a new action-planning loop that pre-verifies preconditions before committing to UI interactions.