Claude Sonnet 5 lands with sharper agent skills — and performance that rivals Opus

Anthropic released Claude Sonnet 5 on Tuesday, calling it the most agent-capable model in its Sonnet line. The model plans tasks, operates a browser, and uses terminal tools autonomously — a meaningful step beyond what Sonnet 4.6 could do.

The new model is available now through Claude Code and the Claude Platform. Developers can call it via API using the model ID “claude-sonnet-5”. Anthropic is offering introductory pricing through August 31, 2026: $2 per million input tokens and $10 per million output tokens. After the promotional period ends, prices rise to $3 and $15 respectively.

On benchmarks, Sonnet 5 shows clear gains over its predecessor. In BrowseComp, an agentic web search evaluation, and OSWorld-Verified, a computer-use benchmark, the improvements are substantial. In some tasks, Sonnet 5’s performance approaches that of Opus 4.8, Anthropic’s flagship model.

Anthropic says Sonnet 5 also improves on safety across the board — lower rates of undesirable behavior compared to Sonnet 4.6, better rejection of malicious requests, stronger resistance to prompt injection attacks, and reduced hallucination and sycophancy.

The release positions Sonnet 5 as a middle-ground option for developers who need strong agentic capabilities without the cost of Opus. With pricing that undercuts many competitors and performance that in some areas closes the gap with Anthropic’s top-tier model, it gives teams building AI agents a practical new option.