THE SHORT VERSION
- Anthropic released Sonnet 5.5 on September 28 for everyday work; on October 7 it followed with Haiku 5.5 for high-volume, cost-sensitive tasks.Anthropic 于 9 月 28 日发布 Sonnet 5.5，面向日常工作；10 月 7 日又推出 Haiku 5.5，面向高频、成本敏感任务。
- Sonnet 5.5’s API input/output rates match Sonnet 5; Anthropic says it generates output over 30% faster and used fewer tokens per task in its tests. On October 7, it also cut Sonnet 5.5 cache-read pricing to $0.10 per million tokens.Sonnet 5.5 的 API 输入和输出单价与 Sonnet 5 相同；Anthropic 称其输出快 30% 以上，并在测试中每项任务用更少 token。10 月 7 日，公司还将 Sonnet 5.5 的缓存读取价格降至每百万 token 0.10 美元。
- Sonnet 5.5 adds cyber safeguards; high-risk requests can fall back to Sonnet 5, while routine software development is intended to remain available.Sonnet 5.5 增加了网络安全防护；遇到高风险请求时可能回退到 Sonnet 5，而常规软件开发预计仍可使用。
Sonnet 5.5 is aimed at the work between quick answers and deep projectsSonnet 5.5 面向快速问答与复杂项目之间的日常工作
Anthropic announced Claude Sonnet 5.5 on September 28 as the second model in its Claude 5.5 family. TechCrunch’s same-day coverage described the release as a cheaper, faster work partner. The company describes it as a faster, lower-cost complement to Opus 5.5. Opus is positioned for complex work that needs sustained judgment; Sonnet is aimed at well-scoped tasks such as fixing bugs, preparing documents, slides, and spreadsheets. At launch, Anthropic said a Haiku model for high-volume work would follow; on October 7, the company released Haiku 5.5, positioning it for high-volume, cost-sensitive tasks such as summaries, subagents, and browser use.Anthropic 于 9 月 28 日发布 Claude Sonnet 5.5，作为 Claude 5.5 系列的第二款模型。TechCrunch 当天的报道将其描述为更便宜、更快的工作伙伴。公司把它定位为 Opus 5.5 的更快、低成本补充。Opus 面向需要持续判断力的复杂工作；Sonnet 更适合边界清晰的任务，例如修复 bug、处理文档、制作演示文稿和电子表格。发布 Sonnet 时，Anthropic 表示会推出面向高频工作的 Haiku；公司在 10 月 7 日发布了 Haiku 5.5，并将其定位为适合摘要、子代理和浏览器使用等高频、成本敏感任务的模型。
That positioning answers a practical question: which model should a team start with for an ordinary task? Sonnet is the middle option in Anthropic’s model range, so a user can reserve Opus for work where broader judgment matters and use Sonnet for a defined review or draft. TechCrunch’s launch coverage similarly describes Sonnet as Anthropic’s mid-tier model and reports that the company emphasized speed and lower cost. Neither source says Sonnet will be the best choice for every workflow.这个定位回答了一个实用问题：日常任务应该先选哪款模型？如果任务是明确定义的审阅或草稿，用户可以从 Sonnet 开始，把需要更广泛判断的工作留给 Opus。TechCrunch 的发布报道也将 Sonnet 描述为 Anthropic 的中档模型，并指出公司在发布时强调速度和成本。两篇材料都没有说 Sonnet 会成为所有工作流程的最佳选择。
Output speed and task cost are separate claims输出速度与任务成本是两个不同的说法
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5. The company lists API prices of $2 per million input tokens and $10 per million output tokens, the same rates as Sonnet 5. It also says Sonnet 5.5 typically needs fewer tokens to do the same work and cost up to 30% less per task in its testing. The difference is between the list price for each token and the total tokens a model uses to finish a task.Anthropic 表示，Sonnet 5.5 的输出速度比 Sonnet 5 快 30% 以上。公司公布的 API 价格为每百万输入 token 2 美元、每百万输出 token 10 美元，与 Sonnet 5 的标价相同。它还表示，Sonnet 5.5 通常用更少 token 完成同一工作，在公司测试中每项任务的成本最多低 30%。这里需要区分每个 token 的标价与模型完成任务时实际使用的 token 总量。
A faster response does not automatically mean a cheaper bill. The bill depends on input length, output length, cached context, tool calls, and how much work the model repeats. Anthropic’s October 7 Haiku 5.5 update cut Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens. A team using cached context should include that later change in its estimate. For someone using a Claude subscription, API prices do not directly describe the plan’s message limits or monthly cost.响应更快并不自动等于账单更低。费用还取决于输入长度、输出长度、缓存上下文、工具调用和模型是否重复工作。Anthropic 在 10 月 7 日的 Haiku 5.5 更新中将 Sonnet 5.5 的缓存读取价格从每百万 token 0.20 美元降至 0.10 美元。使用缓存上下文的团队应把这项后续调整纳入估算。如果使用的是 Claude 订阅服务，API 价格也不能直接说明订阅的消息额度或月费。
Benchmark gains are useful, with a narrow scope基准成绩有用，但范围有限
Anthropic’s performance table reports a large increase over Sonnet 5 on Terminal-Bench 4.0, a command-line agent evaluation, and a smaller gap between Sonnet 5.5 and Opus 5.5 on several knowledge-work tests. The company also says Sonnet 5.5 can approach Opus 5.5 at some effort settings. It explicitly cautions that benchmarks cover only part of a model’s abilities and says Opus remains stronger for complex, open-ended work requiring sustained judgment. For one concrete comparison, Anthropic's table lists 70.6% for Sonnet 5.5 and 10.3% for Sonnet 5 on Terminal-Bench 4.0, under the company's stated test setup. Anthropic notes that GPT-6 Sol was not listed for Terminal-Bench because a public score was unavailable.Anthropic 的性能表显示，Sonnet 5.5 在命令行代理评测 Terminal-Bench 4.0 上高于 Sonnet 5，在多项知识工作评测中则与 Opus 5.5 的差距较小。公司还称，在部分 effort 设置下，Sonnet 5.5 的表现可接近 Opus 5.5。Anthropic 明确提醒，基准只能反映模型能力的一部分，并表示面对需要长期判断的复杂开放任务时，Opus 仍更强。作为一个具体比较，Anthropic 的表格报告 Sonnet 5.5 在 Terminal-Bench 4.0 上为 70.6%，Sonnet 5 为 10.3%，这来自公司所述的测试设置。Anthropic 还说明，由于没有公开分数，表格没有列出 GPT-6 Sol 在 Terminal-Bench 上的成绩。
The footnotes matter. Anthropic notes that an evaluation provider ran some knowledge-work tests on a pre-release deployment that had a structured-output bug, which the company says has since been fixed. That does not make the results unusable, but it changes how a reader should compare scores. TechCrunch’s report describes the model’s speed and lower-cost positioning and notes that Anthropic’s benchmark claims drive part of the launch story. Neither source supplies an independent, broad test of a typical company’s workload.脚注也会影响读者如何解读。Anthropic 说明，某评测机构曾在 Sonnet 5.5 的预发布版本上进行部分知识工作测试；该版本有一个可能影响结构化输出的缺陷，公司称后来已修复。这并不意味着分数不可用，但会影响比较结论。TechCrunch 的报道介绍了模型速度和低成本定位，也指出 Anthropic 的基准说法是发布报道的一部分。两项来源都没有对普通企业工作负载进行广泛独立测试。
Effort settings change the speed and cost tradeoffEffort 设置会改变速度与成本取舍
Anthropic says Claude apps and Claude Code use Medium effort by default, while its developer platform uses High effort by default. Lower effort settings favor faster responses and fewer tokens; higher settings let the model reason longer and check its work more thoroughly. This means a team comparing two Sonnet 5.5 results should record the effort setting, not only the model name. A different setting can change latency and token use even when the prompt stays the same.Anthropic 表示，Claude 应用和 Claude Code 默认使用 Medium effort，而开发者平台默认使用 High。较低设置优先保证响应速度并减少 token 用量；较高设置会让模型进行更长的推理和检查。如果团队比较两次 Sonnet 5.5 结果，应记录 effort 设置，而不能只记模型名称。即使提示完全相同，设置不同也可能改变响应时长和 token 用量。
For a repeatable task, teams can separate quick drafts from decisions that need more checking. A short summary of a named document may suit a lower effort level. A code change that touches several modules may need a higher setting and a human review. Anthropic’s launch materials show that effort is one way to tune the tradeoff, but they do not tell every user what setting will be right for their own project.对重复性任务，团队可以把快速草稿与需要更多核验的决定分开。摘要一份指定文档可能适合较低设置；跨多个模块修改代码则可能需要较高设置和人工审查。Anthropic 的发布资料说明了如何调整速度和成本，但没有替每个团队决定其项目应选哪一种设置。
Cyber safeguards change how some requests are handled网络安全防护会改变部分请求的处理方式
Anthropic says Sonnet 5.5 has stronger cybersecurity capabilities than Sonnet 5 and is the first Sonnet model to launch with safeguards similar to those on Opus 5.5. For higher-risk cybersecurity requests, the system can visibly fall back to Sonnet 5. The company says routine software development and ordinary bug-fixing remain available. Sonnet 5.5 uses the same biology safeguards as Sonnet 5, while a verification program is intended to give qualifying defenders access to a wider range of capabilities.Anthropic 称，Sonnet 5.5 的网络安全能力比 Sonnet 5 更强，也是首款启用类似 Opus 5.5 防护措施的 Sonnet 模型。遇到更高风险的网络安全请求时，系统可能明显回退到 Sonnet 5。公司表示，常规软件开发和普通 bug 修复仍然可用。Sonnet 5.5 使用与 Sonnet 5 相同的生物学防护；面向符合条件的防御人员，公司还计划通过验证计划开放更多能力。
This is a product and policy detail, not a guarantee that the model will classify every request correctly. Developers should expect a visible difference when a request crosses Anthropic’s cyber risk boundary and should check current documentation if their work is defensive security research. For routine coding, the company says the safeguard is designed to leave normal development unaffected. The distinction helps teams plan around a fallback, but does not replace their own authorization and review controls.这是产品与政策细节，并不保证模型能正确识别每个请求。开发者应预期：请求一旦触及 Anthropic 的网络安全风险边界，模型处理方式可能发生变化；如果从事防御性安全研究，应查看当前文档。对于常规编码，公司称防护设计不会影响普通开发。这个区分有助于团队计划回退路线，但不能取代自己的授权与审查措施。
Sonnet 5.5 is available across Anthropic’s listed platformsSonnet 5.5 已在 Anthropic 列出的多个平台提供
Anthropic says Sonnet 5.5 is available on Claude, the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure. The API model identifier is `claude-sonnet-5-5`. The company also says the model can be used with zero data retention, subject to the relevant account and service terms. Developers who move from Sonnet 5 should read the migration guidance; Anthropic notes a changed setting for requests that disable thinking in the API.Anthropic 表示，Sonnet 5.5 已可通过 Claude、Claude API、Amazon Web Services、Google Cloud 和 Microsoft Azure 使用。API 模型标识为 `claude-sonnet-5-5`。公司还称可根据相关账户和服务条款使用零数据保留。开发者从 Sonnet 5 迁移时应阅读迁移指南；Anthropic 特别提到，对关闭思考功能的 API 请求需要使用新的设置项。
A model appearing on several cloud platforms does not make every deployment identical. The available model version, default effort, access controls, logging, and pricing can depend on whether a team uses Anthropic directly or through a cloud provider. Before changing a production route, compare the provider’s documentation and verify how its model identifier, context limits, and safeguards are configured. The launch announcement establishes where Anthropic says the model is available, not that every provider exposes the same settings. The launch page lists its rates at publication, while a cloud partner may show different billing details; check current provider terms before budgeting or sending sensitive material.一个模型出现在多个云平台上，并不意味着每种部署都相同。可用版本、默认 effort、访问控制、日志和价格，可能取决于团队直接使用 Anthropic，还是通过云平台访问。更改生产环境中的模型路由前，应比较对应提供方的文档，并确认其模型标识、上下文限制与防护配置。发布公告说明了 Anthropic 称模型可在哪里使用，却不能证明每个平台的设置完全相同。发布页面列出了推出时的价格，但云服务伙伴的账单细节可能不同；预算或发送敏感材料前，应核对当前提供方的条款。
Compare Sonnet with the work your team actually runs用团队实际运行的工作来比较 Sonnet
A team can make a useful model choice without asking one model to solve an artificial contest. Pick a small set of real, well-scoped tasks: a bug fix with expected tests, a document summary with known source text, or a spreadsheet task with a verified answer. Record the model, effort setting, input context, output quality, tool calls, and time to completion. That gives a more relevant comparison than a single benchmark score or a vendor’s best-case example.团队不必让模型参加一场人为设计的竞赛，也能做出有用选择。可以挑一组真实且边界清楚的任务：带预期测试的 bug 修复、能对照原文的文档摘要，或有标准答案的表格任务。记录模型、effort 设置、输入上下文、输出质量、工具调用次数和完成时间。这样的比较比只看一个基准分数或厂商的最佳示例更贴近团队工作。
Sonnet 5.5 is a candidate when a team wants a capable model for everyday coding and knowledge work and can accept its cyber safeguard behavior. Opus may fit a task that needs broader judgment; Haiku 5.5 is now available for high-volume, cost-sensitive requests such as summaries and subagent work. Compare all three on the task, review cost, and usage pattern rather than assuming Sonnet’s reported savings carry across workloads.如果团队希望在日常编码和知识工作中使用较强模型，并能接受相应的网络安全防护，Sonnet 5.5 可以列入候选。需要更广泛判断的任务可考虑 Opus；Haiku 5.5 已经开放，定位为高频、成本敏感请求，适合摘要和子代理任务。应围绕实际工作、复核成本和用量比较三款模型，不要假设 Sonnet 宣称的节省会在每种负载中出现。
Sources & reporting limits来源与核查范围
Follow Models to find related stories on your next visit.关注模型动态，下次访问时阅读相关文章。
