Claude Sonnet 4.5 is Anthropic's safest AI model yet (www.engadget.com)

Anthropic today unveiled Sonnet 4.5, billing it as the company’s strongest coding model to date after it outperformed its predecessor and pricier siblings on multiple benchmarks — notably setting a record 61.4% on the OSWorld real‑world tasks suite, 17 points ahead of Opus 4.1 — and beating competitors including Google’s Gemini 2.5 Pro and OpenAI’s GPT‑5 on selected tests. The model also substantially extends autonomous agent capabilities, sustaining multi‑step project work for more than 30 hours (vs. ~7 hours for Opus 4 at launch), a key milestone for agentic workflows. API pricing remains $3 per 1M input tokens and $15 per 1M output tokens, and Claude Code gets UX upgrades like a refreshed terminal, checkpoints to snapshot/roll back sessions, in‑conversation file creation, and a Claude for Chrome rollout. Anthropic emphasizes safety: Sonnet 4.5 underwent “extensive” safety training and is claimed to be its least prone model yet to sycophancy, deception, power‑seeking, and encouraging delusional thinking, plus hardened defenses against prompt‑injection. Because of its sophistication, it’s released under Anthropic’s AI Safety Level 3 with filters to block dangerous outputs (e.g., chemical, biological, nuclear). The release tightens Anthropic’s position in enterprise AI (Microsoft has already added Claude models to Copilot 365) and signals continued competition focused on combining stronger coding performance, longer agentic runtimes, and tighter safety guardrails.