Anthropic 发布 Sonnet 5.5:更快更省 token,能力大幅提升Anthropic launches Sonnet 5.5: faster, fewer tokens, major capability jump
Claude Code 负责人 Boris Cherny 展示 Sonnet 5.5 修复 Claude Code bug,速度提升 30%,用量减少 30%。
Claude Code 团队 Cat Wu 表示,用户完成任务量增加约 30%,因为模型更聪明,同样工作需要更少 token。演示中它用 tool calls 清扫院子树叶,比 Sonnet 5 快 24 秒并少用 6K token。
Anthropic 研究员 Alex Albert 称 Sonnet 5.5 保留了 Opus 5.5 那种清晰写作和极速响应的感觉,是相对 Sonnet 5 的重大能力跃升,非常适合迭代。
Box CEO Aaron Levie 分享早期评测:在复杂企业知识工作评估中整体提升 4 分,完成交付物速度约 2.4 倍,token 少 12%。金融服务业 +18 分、法律 +7、生命科学 +8、公共部门 +7。具体案例包括尽职调查抓住算术错误、拒绝编造租约标准、正确计算样本标准差等。客户很快就能在 Box AI Studio 用 Sonnet 5.5 构建 agent。
Every CEO Dan Shipper 测试后认为写作能力大幅提升(甚至在修订任务上超越 Opus 5.5),速度和成本优于 Opus 5.5,成为团队快速迭代编码和设计的首选。
Claude Code 团队 Cat Wu 表示,用户完成任务量增加约 30%,因为模型更聪明,同样工作需要更少 token。演示中它用 tool calls 清扫院子树叶,比 Sonnet 5 快 24 秒并少用 6K token。
Anthropic 研究员 Alex Albert 称 Sonnet 5.5 保留了 Opus 5.5 那种清晰写作和极速响应的感觉,是相对 Sonnet 5 的重大能力跃升,非常适合迭代。
Box CEO Aaron Levie 分享早期评测:在复杂企业知识工作评估中整体提升 4 分,完成交付物速度约 2.4 倍,token 少 12%。金融服务业 +18 分、法律 +7、生命科学 +8、公共部门 +7。具体案例包括尽职调查抓住算术错误、拒绝编造租约标准、正确计算样本标准差等。客户很快就能在 Box AI Studio 用 Sonnet 5.5 构建 agent。
Every CEO Dan Shipper 测试后认为写作能力大幅提升(甚至在修订任务上超越 Opus 5.5),速度和成本优于 Opus 5.5,成为团队快速迭代编码和设计的首选。
Claude Code lead Boris Cherny showed Sonnet 5.5 fixing a Claude Code bug while running 30% faster and using 30% fewer tokens.
Claude Code team member Cat Wu reported users complete ~30% more tasks because the model is smarter and needs fewer tokens for the same work. In a demo it raked a yard of leaves via tool calls, finishing 24s faster than Sonnet 5 with 6K fewer tokens.
Anthropic researcher Alex Albert said Sonnet 5.5 retains the clear writing and speed he liked in Opus 5.5, delivering a major capabilities jump over Sonnet 5 and making it excellent for iteration.
Box CEO Aaron Levie shared early-access results on complex enterprise knowledge-work evals: overall +4 points, roughly 2.4× faster to finished deliverables on 12% fewer tokens. Vertical gains included +18 pts in Financial Services, +7 in Legal, +8 in Life Sciences, and +7 in Public Sector. Specific wins: catching miscalculated interest and mispriced options in a deal book, refusing to invent lease benchmarks, correctly computing sample standard deviations, and pulling accurate progress-monitoring figures—all completed far faster. Customers will soon be able to build agents with Sonnet 5.5 in Box AI Studio.
Every CEO Dan Shipper found dramatically improved writing (beating Opus 5.5 on revision tasks), plus speed and cost advantages that make it the preferred model for quick iterative coding and design work.
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →
Claude Code team member Cat Wu reported users complete ~30% more tasks because the model is smarter and needs fewer tokens for the same work. In a demo it raked a yard of leaves via tool calls, finishing 24s faster than Sonnet 5 with 6K fewer tokens.
Anthropic researcher Alex Albert said Sonnet 5.5 retains the clear writing and speed he liked in Opus 5.5, delivering a major capabilities jump over Sonnet 5 and making it excellent for iteration.
Box CEO Aaron Levie shared early-access results on complex enterprise knowledge-work evals: overall +4 points, roughly 2.4× faster to finished deliverables on 12% fewer tokens. Vertical gains included +18 pts in Financial Services, +7 in Legal, +8 in Life Sciences, and +7 in Public Sector. Specific wins: catching miscalculated interest and mispriced options in a deal book, refusing to invent lease benchmarks, correctly computing sample standard deviations, and pulling accurate progress-monitoring figures—all completed far faster. Customers will soon be able to build agents with Sonnet 5.5 in Box AI Studio.
Every CEO Dan Shipper found dramatically improved writing (beating Opus 5.5 on revision tasks), plus speed and cost advantages that make it the preferred model for quick iterative coding and design work.