Anthropic is rolling out Claude Sonnet 5, its latest mid-tier model with a heavy emphasis on agentic capabilities — and a subtle price hike disguised as a launch deal. The new model is designed for autonomous planning, browser control, and terminal tool use, and it significantly outperforms its predecessor Sonnet 4.6 on benchmarks like BrowseComp (intelligent search) and OSWorld-Verified (computer operations). By tweaking an "effort" parameter, users can push its performance close to Anthropic's flagship Opus 4.8.
On the safety front, Sonnet 5 boasts a lower hallucination rate and improved defenses against malicious requests and prompt injection attacks. It's not specifically trained for cybersecurity — that remains Opus's domain — but it defaults to real-time network security protection. The catch? Anthropic upgraded the tokenizer, which now produces 1.0 to 1.35 times more tokens for the same text. To soften the blow, the company is offering launch pricing: $2 per million input tokens and $10 per million output tokens (valid until August 31, 2026, after which prices revert to $3 and $15). That's meant to keep the switch from Sonnet 4.6 cost-neutral.
Sonnet 5 is available now to all users across Free, Pro, Max, Team, and Enterprise plans, and it works inside Claude Code.