Claude Sonnet5 releases: Intelligent agent capabilities approaching flagship, first release discounts to hedge token inflation According to Beating monitoring, Anthropic has officially launched its mid-range flagship model, Claude Sonnet 5, which features powerful Agent autonomous planning, browser and terminal tool calling, and other automatic running capabilities. The new model significantly outperforms the previous generation Sonnet 4.6 in evaluations such as BrowseComp (intelligent search) and OSWorld Verified (computer operation). Users can adjust the effort level to achieve performance comparable to the flagship model Opus 4.8. In terms of security performance, the new model has a lower illusion rate compared to Sonnet 4.6, and has improved defense against malicious requests and prompt injection attacks. Although the model has not been specifically trained for network security and its security capabilities are much lower than Opus 4.8, real-time network security protection mechanisms are enabled by default. In addition, the model has upgraded the tokenizer, and the number of tokens generated for the same text has increased by 1.0 to 1.35 times compared to the old version. Therefore, the initial discount price is set at $2 per million tokens for input and $10 for output (until August 31, 2026, after which the standard prices of $3 and $15 will be restored), aiming to keep the actual cost of users switching models basically the same. At present, the model is open to all users of Free, Pro, Max, Team, and Enterprise, and can be used in Claude Code. [Original link]