Claude Sonnet 5: More Agentic, Lower Cost

The agentic AI era has seen rapid advances in large, expensive Opus-class models, leaving Sonnet-class models struggling to keep up. Developers who rely on affordable, capable models for tool use, coding, and autonomous tasks faced a trade-off between cost and performance. Claude Sonnet 5 directly addresses this tension by narrowing the gap with Opus 4.8 while maintaining lower prices, specifically $2 per million input tokens and $10 per million output tokens through August 2026.

Sonnet 5 is built to be the most agentic Sonnet model yet, with significant improvements in reasoning, tool use, coding, and knowledge work over its predecessor Sonnet 4.6. Early access partners report it finishes complex multi-step tasks that previously stalled, checks its own output unprompted, and handles messy technical contexts like brownfield code and legacy systems. Safety evaluations show lower rates of undesirable behaviors than Sonnet 4.6, including reduced hallucination and sycophancy, though it scores slightly worse than Opus 4.8 on some misalignment audits. Cyber safeguards are enabled by default, as the model shows a slightly higher partial success rate on cybersecurity exploit tasks than Sonnet 4.6.

The takeaway for builders is clear: Claude Sonnet 5 offers a compelling price-performance point for agentic workflows. It is available across all plans, in Claude Code, and via the Claude API with introductory pricing that makes the transition from Sonnet 4.6 roughly cost-neutral. Developers can adjust effort levels to balance cost and accuracy, making Sonnet 5 a practical choice for day-to-day automation, software engineering, legal research, and other agentic tasks where follow-through and technical grounding matter.