Today, Anthropic officially released Claude Sonnet 5, touted as the most agentic-capable model in the Sonnet series to date. It can plan, utilize tools such as browsers and terminals, and autonomously execute tasks—capabilities that several months ago required larger, more expensive models. Sonnet 5 significantly narrows the gap with the flagship Opus 4.8 model. Compared to its predecessor, Sonnet 4.6, it shows substantial improvements across key agentic metrics including reasoning, tool calling, coding, and knowledge work. Safety evaluations indicate an overall lower rate of undesirable behavior than Sonnet 4.6, though slightly higher rates were observed during automated behavioral audits compared to Opus 4.8 and Mythos Preview. Effective immediately, Sonnet 5 is live on all subscription plans: Free and Pro users receive access by default, while Max, Team, and Enterprise customers may opt-in. The model is also now available via Claude Code and the Claude Platform under the identifier claude-sonnet-5. Pricing features a promotional rate valid until August 31, 2026: input at $2 per million tokens and output at $10 per million tokens. After this period, pricing will adjust to input at $3 per million tokens and output at $15 per million tokens. Anthropic did not specifically train Sonnet 5 against cybersecurity threats. Evaluations show its capability to perform dangerous tasks like developing software exploits is far below that of Opus 4.8 and Mythos 5. Cybersecurity protections are enabled by default, allowing real-time detection and blocking of hazardous network usage behaviors—a protection level equivalent to Opus 4.7/4.8 but less stringent than that applied to Mythos 5. Sonnet 5 outperforms Sonnet 4.6 in hallucination reduction (by 45%), resistance to sycophancy, and defense against prompt injection attacks. Users can select varying levels of reasoning effort based on task complexity to flexibly balance cost and performance.