Claude Sonnet 5.5 发布：编码性能和 Opus 5.5不相上下 价格是后者一半
同价升级的重点，不是单个基准分数，而是修 bug、操作工具和制作文档时少等待、少迭代；结合官方演示看清它能做什么。
Anthropic 发布了 Claude Sonnet 5.5：Sonnet 5 的同价升级，也是 Opus 5.5 更快、价格更低的搭档。它要解决的痛点很实际——修一个 bug、改一组文件或做一份汇报，模型也许都能完成，但如果反复试错、调用工具和生成长篇中间结果，等待时间与整项任务的费用就会上升。Sonnet 5.5 希望让这些边界清楚的日常工作更快交付，同时把需要长期判断的开放任务留给 Opus。
这次发布的主要特点可以先记住四点：
- 更快完成同一项任务。 Sonnet 5.5 与 Sonnet 5 的 API 单价相同；Anthropic 称它输出快 30% 以上、通常使用更少 token，在其测试中每任务成本最多低 30%。
- 更会连续做编码工作。 提升不只体现在补代码，还体现在命令行操作、多文件修改、工具调用和反复迭代。
- 更接近可交付的办公产物。 它能按材料和模板制作文档、幻灯片与表格，知识工作评测的部分得分已接近 Opus 5.5。
- 可以按任务调节投入。 用 effort 档位在速度、费用和质量之间取舍；简单任务不必一律开到 Max。
下面按这四点展开：它到底快在哪里、能接哪些工作，以及怎么设置才划算。
同样的价格，为什么一件事做下来可能更省
先看两个容易混淆的价格。Sonnet 5.5 输入每百万 token 2 美元，输出 10 美元，缓存读取 0.20 美元、缓存写入 2.50 美元；Sonnet 5 也是同一档价格。与 Opus 5.5 相比，它的输入、输出和缓存写入单价是后者的一半，缓存读取则同为 0.20 美元。
| API 计费项（每百万 token） | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| 输入 | 2 美元 | 4 美元 |
| 输出 | 10 美元 | 20 美元 |
| 缓存读取 | 0.20 美元 | 0.20 美元 |
| 缓存写入 | 2.50 美元 | 5 美元 |
但只比单价会漏掉更重要的一半。假设两个模型完成同一个修复，其中一个反复搜索、重写和调用工具，最终仍可能花得更多。Anthropic 的成本—得分图因此把横轴设为每次任务的实际成本，纵轴设为得分；同一个模型的 Low、Medium、High、Xhigh、Max effort 分别是不同的点。越靠左上，代表在这项评测里花更少的钱拿到更高的分数。
在几项评测中，Sonnet 5.5 用 Low 或 Medium effort 就超过了 Sonnet 5 的最佳成绩，而每任务成本约为后者的十分之一。到了高 effort，它有时接近 Opus 5.5 的得分，费用也逐渐接近。也就是说，升级的价值不只是“换个更聪明的模型”，而是让常规任务用较少投入达到可用质量；十分之一是这些评测中的结果，不是所有账单都会打这个折。
编码：从补一行代码，进到连续操作和完成任务
Sonnet 5.5 最醒目的数字是 Terminal-Bench 4.0 的 70.6%，Sonnet 5 为 10.3%。这项评测要求代理在命令行里完成多步骤专业任务，所以它反映的不是孤立地补一行代码，而是读环境、操作工具、修改、运行和收尾这一串动作。发布表中 Opus 5.5 为 66.4%；其脚注明确说这个 Opus 数字取自 Xhigh effort。Sonnet 在这个基准上领先，不等于它在所有复杂编码项目里都比 Opus 强。
其他两个编码基准对应不同工作。CursorBench 4.0 取自真实 Cursor 会话中的多文件、含糊任务，Sonnet 5.5 为 55.5%，Sonnet 5 为 34.1%，Opus 5.5 为 57.8%。FrontierCode 看代码变更能否不经人工修改就合并；Sonnet 5.5 的 Xhigh 结果是 52.1%，Opus 5.5 为 54.4%。三项测试合起来说明，它在读环境、使用工具、修改并交付变更这一整串工作上进步明显。
真实产品测试更能说明这些能力如何用起来。Base44 在 118 次应用构建中记录到，Sonnet 5.5 的应用得分与 Opus 5 持平，平均只需 3.6 轮迭代，Opus 5 为 7.7 轮。Unity 的测试会重新打开项目并在运行时检查结果；它报告 Sonnet 5.5 完成了多步骤编辑器与编码任务中的 90%。Epic Games 提到游戏系统架构和多小时任务，CodeRabbit 则准备先迁移简单、中等难度的代码审查。这些是各家公司自己的测试和使用计划，不是同一张排行榜。
ClaudeDevs 的短视频给出另一个更小的例子：一个调度器让四个代理从不同位置并行尝试解谜。视频标注的是 10 次运行的中位数，Sonnet 5.5 在 10 秒、约 2.3 万 token、0.11 美元时完成；Sonnet 5 在 15 秒、约 2.8 万 token、0.15 美元时完成。差异不只在模型回答速度，还在整条工具调用链里走了多少步。
文档、幻灯片和表格：从找材料走向成品
Anthropic 把 Sonnet 5.5 的另一重点放在知识工作：不是只回答“财报里有什么”，而是把材料读完、组织要点，再按指定格式做出可用文件。GDPval-AA 涵盖 44 种职业、九个行业的工作任务，Sonnet 5.5 得到 1844，Opus 5.5 为 1846，Sonnet 5 为 1449。长流程知识工作评测 AA-Briefcase v1.1 中，三者分别是 1811、1822 和 1359。两项分数显示，Sonnet 5.5 在这些任务上已非常接近 Opus 5.5。
发布页给了一个更直观的例子：把一家上市公司的季度财报材料、电话会记录和幻灯片模板交给模型，要求制作 10 页经营复盘；Anthropic 请的两位专家认为第一稿就能发送。这里的重点不是“会写幻灯片”，而是它能把找材料、提炼重点、遵循模板连成一项任务。Claude 官方主题帖还放出了沙丘网页的并列演示，让人看到它在视觉产物上的具体表现。
实际团队看重的还包括“同样质量用了几步”。Slack 说未改提示词时，其离线 Slackbot 测试几乎全面优于 Sonnet 5，输出 token 少约 14%；Zendesk 称在数百个客服回复和升级案例中，工单处理快了 20%；Balyasny Asset Management 说，在其 2441 道金融任务的私有测试里，每答案约用 12.1 万 token，而 Sonnet 5 约用 49.7 万；Lovable 的编码评测则记录到工具调用少约三分之一、shell 运行约减半。Box 报告 2.4 倍的处理速度和少 12% 的总 token；Atlassian 预计 Rovo Agents 可提速最多 30%。这些数字来自不同公司的不同任务，意义是提示团队应测自己的“每件事成本”，不能加总或相互当作复现。
其他任务也有提升：Sonnet 5.5 在 Humanity’s Last Exam 得到 64.5%（使用工具），OSWorld 2.1 得到 80.1%（部分任务集），Chartography 看图表得到 61.6%（不使用工具）；Sonnet 5 对应为 54.9%、57.0% 和 15.6%。这些分数指向更广的推理、电脑操作和读图能力，但它们不是同一种测试，不能相加成一个总分。
effort：先按任务调档，不要习惯性开 Max
effort 是模型在一项任务中投入多少推理与检查。Claude Code 和 Claude 应用默认 Medium，Claude Platform 默认 High。低档通常更快、更省；高档可以多检查，适合更难的任务，但投入增加也会拉高每任务费用。官方四张成本—得分图把这件事画得很清楚：蓝色的 Sonnet 5.5 在较低档位就能超过旧版不少结果；档位继续升高，得分和成本都向 Opus 靠拢。
FrontierCode 还出现一个值得记住的例外：Sonnet 5.5 在 Xhigh 得到 52.1%，开到 Max 反而是 46.2%。Anthropic 解释，Max 更常触发多代理代码审查；两个被检查的案例因此超时或改了任务范围外的代码。对修 bug、改文件这类有明确验收标准的工作，先从 Medium 或较低档位测起，结果不够再加投入，比一上来开 Max 更符合这次发布展示的成本优势。
查看四张官方成本—得分图：知识工作、CursorBench、FrontierCode、Terminal-Bench
同一提示词，两代模型的网页作品有什么不同
发布页让 Sonnet 5 与 5.5 各用一个 HTML 文件完成三项可视化任务。静态图方便看清两边到达了哪一步；下方的原始动态程序还能看到代码生成和作品运行的过程。
400 只椋鸟群飞
Sonnet 5 的截图仍在写入代码（4520 token）；Sonnet 5.5 在输出 4158 token 后，画面已有运行中的群飞效果。
风塑造沙丘
旧版截图仍在写代码（2966 token）；新版在输出 2720 token 后，呈现出多层线条沙丘与红日。
24 个小钟组成大钟
这一组的差别尤其直观：Sonnet 5 在写入 27,581 token 时仍停留在代码画面；Sonnet 5.5 输出 14,386 token 后，画面上的小钟指针已拼出大数字。
想看完整过程，可展开下面六个原始动态程序。旧版截图拍在编写过程中，并不意味着它最终做不出作品；这里展示的是这些官方演示在对应时刻的进度。
播放三组对照的六个原始动态程序
椋鸟群飞 · Sonnet 5
椋鸟群飞 · Sonnet 5.5
风塑造沙丘 · Sonnet 5
风塑造沙丘 · Sonnet 5.5
小钟组成大钟 · Sonnet 5
小钟组成大钟 · Sonnet 5.5
上线使用时，还会遇到哪些设置和限制
除了性能，Sonnet 5.5 还调整了安全处理。Anthropic 的约 1850 个自动化行为场景测试中，它在多数对齐、诚实和抗滥用指标上追平或优于 Sonnet 5；Opus 5.5 的整体结果仍略好。这说明新版通过了更严格的发布检查，但具体工作流仍需要照常设置权限和验收。
实际会遇到的变化有三类：普通找 bug、修 bug 不受网络安全回退影响，更高风险的网络安全请求则可能转到 Sonnet 5；生物领域沿用原有防护，少数正当请求可能被误拦；跨账号移动对话时，扩展后的 preserved thinking 会要求在新账号重新生成推理内容。对一般写代码和做文档的人，前两项可能不常出现；团队如果有安全研究或跨组织会话迁移，就需要提前测试。
哪些任务先换 Sonnet 5.5
先从输入、验收标准都清楚的任务试起：修一个可复现的 bug、按模板做经营复盘、批量处理客服工单。让 Sonnet 5.5 在 Medium 或较低 effort 跑一组真实样本，记录完成率、返工时间、工具调用、耗时和最终费用，再与 Sonnet 5 或 Opus 比。若目标不断变化、需要长期取舍，Opus 5.5 仍是 Anthropic 推荐的起点。
开发者可在 Claude Platform 使用模型 ID claude-sonnet-5-5，也可通过 AWS、Google Cloud 和 Microsoft Azure 使用；Anthropic 称它支持零数据保留。已经把 Sonnet 的 thinking 关闭的团队，迁移到 5.5 时还需按官方迁移指南改用 between_tools 设置。下一款面向高吞吐、成本敏感应用的 Haiku 5.5，发布页说将在随后加入家族；这里还没有可据以选型的实测数据。
Sonnet 5.5 的价值不是每个榜单都赢，而是让更多日常编码和办公任务不必动用 Opus，也能较快做出可交付的结果。最值得验证的不是单个分数，而是你自己的工作究竟少等了多久、少返工了几轮、少花了多少钱。
Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work.
Sonnet 5.5 is a faster, lower-cost complement to Claude Opus 5.5. Where Opus 5.5 is built for complex work requiring careful judgment, Sonnet 5.5 is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets. It’s also got a sharp eye for design. Claude Haiku 5.5, built for high-volume and cost-sensitive applications, will join the Claude 5.5 family in the coming weeks.
Sonnet 5.5 improves over Sonnet 5 on:
Performance. Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, an agentic coding evaluation, compared to Sonnet 5’s 10.3%. It scores two points below Opus 5.5 on GDPval-AA, a test of real-world work across a variety of occupations. And it’s strong on long-horizon work and image understanding—it’s the first Sonnet model to beat Pokémon Red working only from screenshots.
Collaboration. Like Opus 5.5, Sonnet 5.5 writes more clearly than our previous generation of models; early testers described it as a better partner for collaboration than Sonnet 5. Its speed also makes it well suited to fast iteration on less complex tasks.
Cost. Sonnet 5.5 is priced the same as Sonnet 5 at $2 per million input tokens, $10 per million output tokens, and $0.20 per million tokens for cache reads, but it typically needs far fewer tokens to do the same work. In our testing, it costs up to 30% less per task than its predecessor.
Speed. Sonnet 5.5 generates outputs 30%+ faster than Sonnet 5, making it our fastest Sonnet model to date.
Alignment and safety. On our automated behavioral audit, Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment. Because its cybersecurity capabilities are comparable to Opus 5’s, it’s the first Sonnet model to launch with cyber safeguards and fallbacks like those we’ve developed for our most capable models. Its biology safeguards are the same as Sonnet 5’s. Both safeguards target a narrow set of high-risk requests; routine software development and most life sciences work are unaffected.
Performance
Sonnet 5.5 improves on Sonnet 5 across domains—in some cases dramatically. On several evaluations, Sonnet 5.5 at Max effort even performs comparably to Opus 5.5. However, benchmark scores capture only one facet of a model’s capabilities; in our own testing, and in that of external testers, Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment.
| Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol | |
|---|---|---|---|---|
| Agentic codingTerminal-Bench 4.0 | ||||
| --- | --- | --- | --- | |
| Agentic codingTerminal-Bench 4.0 | 70.6% | 10.3% | 66.4%¹ | — |
| Agentic codingFrontierCode 1.1 (Main) | ||||
| Agentic codingFrontierCode 1.1 (Main) | 46.2%Max² | 42.4% | 54.4% | 49.3% |
| 52.1%Xhigh | ||||
| Agentic codingCursorBench 4.0 | ||||
| Agentic codingCursorBench 4.0 | 55.5% | 34.1% | 57.8% | — |
| Knowledge workGDPval-AA v2.1³ | ||||
| Knowledge workGDPval-AA v2.1³ | 1844 | 1449 | 1846 | 1487⁴ |
| Knowledge workAA-Briefcase v1.1³ | ||||
| Knowledge workAA-Briefcase v1.1³ | 1811 | 1359 | 1822 | 1483⁴ |
| Multidisciplinary reasoningHumanity’s Last Exam | ||||
| Multidisciplinary reasoningHumanity’s Last Exam | 64.5%with tools | 54.9%with tools | 67.7%with tools | — |
| Computer useOSWorld 2.1 | ||||
| Computer useOSWorld 2.1 | 80.1%partial | 57.0%partial | 81.8%partial | — |
| Visual chart recognitionChartography | ||||
| Visual chart recognitionChartography | 61.6%no tools | 15.6%no tools | 64.4%no tools | 53.6%⁴no tools |
For details on how we run our evaluations, see the Sonnet 5.5 System Card.
The charts below plot each model’s score against its cost per task at every effort level. As effort goes up, models typically work for longer, leading to a higher cost per task but generally also a higher score. The closer a point is to the top left of the chart, the more capability it delivers per dollar.
On several benchmarks, Sonnet 5.5 at Low or Medium effort beats Sonnet 5’s best score for about a tenth of the cost per task. It complements Opus 5.5 best when running at lower effort settings, where it costs less per task. At higher settings, it can perform comparably at a similar cost.
Agentic terminal codingAgentic coding: FrontierCodeAgentic coding: CursorBenchKnowledge work: AA-Briefcase
Agentic terminal codingAgentic coding: FrontierCodeAgentic coding: CursorBenchKnowledge work: AA-Briefcase
Terminal-Bench 4.0Accuracy vs. cost
- Sonnet 5.5
- Opus 5.5
- Sonnet 5
- GPT-5.6 Sol
010203040506070Score (%)12510Cost per attempt (USD, log scale)LowMedHighXhighMax
Terminal-Bench 4.0 measures how well a model can complete complex, multi-step professional tasks within a command-line interface. At Medium effort, the default in the Claude apps, Sonnet 5.5 far exceeds Sonnet 5’s best score for less than a tenth of the cost per task.
Terminal-Bench and OpenAI did not report GPT-6 Sol performance publicly, so we report GPT-5.6 Sol here.
FrontierCode v1.1, main setAccuracy vs. cost
- Sonnet 5.5
- Opus 5.5
- Sonnet 5
- GPT-6 Sol
3035404550550Score (%)0.250.501251020Cost per task (USD, log scale)LowMedHighXhighMax
FrontierCode measures whether an agent’s code changes would be merged. At High effort, the default on the Claude Platform, Sonnet 5.5 matches GPT-6 Sol’s best score for about a fifth of the cost per task.²
CursorBench 4.0Accuracy vs. cost
- Sonnet 5.5
- Opus 5.5
- Sonnet 5
- GPT-5.6 Sol
20304050600Score (%)0.5012510Cost per task (USD, log scale)LowMedHighXhighMax
CursorBench evaluates coding agents on ambiguous, multi-file tasks taken from real Cursor sessions. Sonnet 5.5 at Low effort exceeds Sonnet 5’s best score for less than a tenth of the cost per task.
CursorBench 4.0 does not report GPT-6 Sol performance publicly, so we report GPT-5.6 Sol here.
AA-Briefcase v1.1Accuracy vs. cost
- Sonnet 5.5
- Opus 5.5
- Sonnet 5
- GPT-6 Sol
900110013001500170019000Elo0.100.200.501251020Cost per task (USD, log scale)LowMedHighXhighMax
On AA-Briefcase, a new benchmark of long-horizon knowledge work, Sonnet 5.5 at Medium effort bests Sonnet 5’s best score for about one ninth of the cost per task.³
Coding
Sonnet 5.5’s jump in performance is particularly noticeable in coding. At High effort on FrontierCode, it scores 10 points higher than Sonnet 5 at the same setting, at about one fifteenth of the cost per task. On CursorBench, which tests models on tasks from real Cursor coding sessions, its best score is within about two points of Opus 5.5.
Early testers appreciated how quickly Sonnet 5.5 can understand a codebase. They were also struck by its efficiency: in head-to-head runs, it batched tool calls together more than Sonnet 5, leading to fewer steps and lower costs.
Epic GamesEveryCodeRabbitSpaceXAIBase44UnityCreator
Epic GamesEveryCodeRabbitSpaceXAIBase44UnityCreator
“In Epic’s early testing, Claude Sonnet 5.5 cleared the same quality bar you’d expect from a higher-tier model, holding up on a system design audit and a data flow review. The new model managed tens of thousands of lines of code for gameplay system architecture, kept responses snappy, handled multi-hour tasks, and delivered with less prescriptive prompting.”
CompanyEpic Games
AuthorDaniel Vogel, Chief Operating Officer
“Claude Sonnet 5.5 cooks. Fast at coding and can be steered quickly in iterative workflows. But it can still work long if it needs to. It’s got some of Opus 5.5’s natural writing upgrades, which makes it more fun to work with.”
CompanyEvery
AuthorTyler Nishida, Designer
“Claude Sonnet 5.5 shows better judgment than Sonnet 5 across different levels of complexity, while spending significantly fewer output tokens. Sonnet 5’s tendency to reach for web search too often and its high token use are both gone in this new model. We plan to move simple and moderate reviews over now, and more in the coming weeks.”
CompanyCodeRabbit
AuthorDavid Loker, VP of AI
“Claude Sonnet 5.5 delivers frontier-level performance on CursorBench 4.0 at 55.5%, second only to Opus 5.5. We think it will be a hit with developers looking to balance performance with cost.”
CompanySpaceXAI
AuthorSualeh Asif, Director of ML
“Across 118 real app builds, Claude Sonnet 5.5 produced apps that scored level with Opus 5. It got there in 3.6 iterations per build on average, where Opus 5 took 7.7. It had the fewest failed tool calls of any model we compared. It also rarely stopped mid-build to ask the user a question, so fewer builds stall waiting on someone to answer.”
CompanyBase44
AuthorGabriel Grinberg, AI Engineering Lead
“At Unity, we have a high bar for task completion. Projects are reopened and results are checked at runtime, so a task only counts when the change works, not when the model says it’s done. The majority of Claude Sonnet 5.5’s work passed that check. It also completed 90% of tasks in our multi-step Unity Editor and coding benchmark, beating similar models.”
CompanyUnity
AuthorSam Zhang, Creative Technologist
“When Claude Opus 5.5 sets the architecture and general framework for a game, I would feel confident in letting Sonnet 5.5 implement it. I’m impressed with Sonnet 5.5’s handling of long-running, complex tasks.”
CompanyCreator
AuthorKevin Ngo, Creative Coder
Knowledge work
Sonnet 5.5 shows gains in multiple areas of knowledge work. On GDPval-AA, which tests models on real-world tasks across 44 occupations and nine major industries, Sonnet 5.5 scores nearly level with Opus 5.5 and about 400 points above Sonnet 5. It’s close to Opus 5.5 in computer use and chart recognition, and clearly outperforms Sonnet 5 and GPT-6 Sol on long-horizon knowledge work.
Early testers highlighted less quantifiable improvements. They found it to be a more natural conversational partner and remarked on its knack for design, noting that it adds polish to user interfaces and can follow slide templates to create decks that require minimal editing. In one internal test, we gave it a public company’s quarterly earnings materials and call transcripts, along with a slide template, and asked for a 10-slide operating review. Two experts judged its first draft to be ready to send as is.
SlackZendeskBalyasny Asset ManagementBoxLovableAtlassian
SlackZendeskBalyasny Asset ManagementBoxLovableAtlassian
“Without changing any of our prompts, Claude Sonnet 5.5 did better than Sonnet 5 on almost all of our offline Slackbot evals, in fewer steps and with about 14% fewer output tokens. When someone gives Slackbot a task, quality and speed are what matter most, and Sonnet 5.5 allows Slackbot to deliver better outcomes for users, faster.”
CompanySlack
AuthorCurtis Allen, Principal Engineer
“We fed Claude Sonnet 5.5 hundreds of real support use cases across replies and escalation requests. It made fewer wrong decisions and resolved tickets faster than the Claude models we use in production today. Tickets were processed 20% faster, getting our customers the help they need without the wait.”
CompanyZendesk
AuthorAbhinay Kathuria, Director of AI
“On our private suite of 2,441 finance tasks covering Q&A, extraction, analysis, and forecasting, Claude Sonnet 5.5 scored ahead of Sonnet 5 and used about 121k tokens per answer, where Sonnet 5 used 497k. On our analyst search and retrieval work, it was better than Sonnet 5 in almost every way. For high-volume workflows, it had the best quality-to-cost tradeoff of the seven models we ran.”
CompanyBalyasny Asset Management
AuthorJoe Poirier, Senior AI Engineer
“Claude Sonnet 5.5 will give our customers in financial services and healthcare the confidence to use it for their most sensitive work. Sonnet 5.5 rechecks data in source documents, catching errors that Sonnet 5 failed to spot. Compared to the last model, Sonnet 5.5 was more accurate, 2.4x faster, and used 12% fewer total tokens.”
CompanyBox
AuthorYashodha Bhavnani, VP of AI Products
“Claude Sonnet 5.5 thinks in fewer, more robust steps, so builders wait less to see progress. Our coding evals showed a third fewer tool calls and roughly half the shell runs to finish a task. For everyday coding and higher-effort conversations, that means faster iteration and a smoother build loop.”
CompanyLovable
AuthorFabian Hedin, Co-founder and CTO
“With millions of Rovo-assisted actions powering our customers’ workflows each month, execution speed is critical. Claude Sonnet 5.5 will allow teams to run their Rovo Agents up to 30% faster than they could with Sonnet 5. I am excited to offer customers this choice.”
CompanyAtlassian
AuthorJamil Valliani, Head of Product, AI
Cost and speed
Pricing
| Price per 1M tokens | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|
| Cache reads | $0.20 | $0.20 |
| Cache writes | $2.50 | $5 |
| Input tokens | $2 | $4 |
| Output tokens | $10 | $20 |
Sonnet 5.5 requires fewer tokens per task than Sonnet 5, so it’s less expensive to run. It also generates output 30%+ faster, and its efficiency is immediately noticeable:
Creating murmurationsShaping sand dunesBuilding a clock of clocks
Sonnet 5Sonnet 5.5
Prompt:
A murmuration of 400 starlings in one HTML file
Sonnet 5Sonnet 5.5
Claude Sonnet 5
Claude Sonnet 5.5
Prompt:
Wind shaping sand dunes in one HTML file
Sonnet 5Sonnet 5.5
Claude Sonnet 5
Claude Sonnet 5.5
Prompt:
A clock made of 24 small clocks in one HTML file
Sonnet 5Sonnet 5.5
Claude Sonnet 5
Claude Sonnet 5.5
Sonnet 5Sonnet 5.5
Adjusting the effort level lets you balance cost and speed against overall quality. In Claude Code and our apps, the default effort is set to Medium, while the Claude Platform defaults to High. At lower settings, Claude answers faster and uses fewer tokens, which suits routine work. At higher settings, Claude reasons for longer and checks its work more thoroughly.
Safety
Alignment
Sonnet 5.5 doesn’t advance the frontier of our models’ capabilities, so our alignment assessment focused on a targeted set of risks that apply to models of any capability level, including acting against users’ interests, misleading users, and cooperating with high-stakes misuse.
On our automated behavioral audit, which tests Claude across roughly 1,850 scenarios, Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment, resistance to misuse, and honesty. On our newer containment evaluations, Sonnet 5.5 comes close to Opus 5.5, the best model we tested, in how rarely it tries to escape its sandbox, and it’s the least likely of any of our models to probe the limits of its containers. Across the full audit, Opus 5.5 still performs slightly better overall, but we found no evidence that Sonnet 5.5 pursues goals that conflict with the user’s intention.
As we described in our recent alignment assessment, no set of evaluations reliably catches every failure, and Sonnet 5.5 may have tendencies we haven’t found, which is why we pair our own alignment work with the safeguards described below.
Safeguards
Cybersecurity. Sonnet 5.5’s cyber capabilities are a large improvement over Sonnet 5’s, so we’re deploying it with safeguards similar to those on Opus 5.5. Users can still find and fix bugs in their code as part of routine software development, but higher-risk cybersecurity tasks will visibly fall back to Sonnet 5. Soon, cyberdefenders will be able to apply to our expanded Cyber Verification Program for tiered access to more advanced capabilities on Sonnet 5.5, Opus 5.5, and Claude Mythos models.
Biology. Sonnet 5.5 uses the same set of biology safeguards as Sonnet 5. These target harmful requests; most research, education, and clinical work is unaffected, though some microbiology and virology requests may be flagged in error. Organizations can apply to our Life Sciences Verification Program for access to safeguards designed for the full breadth of biology-related work.
Distillation. Distillation attacks, in which attackers use thousands of fake accounts to extract a model’s capabilities at industrial scale, allow bad actors to create highly capable models without the safeguards we build into Claude. Because Sonnet 5.5 is far more capable than its predecessor, it’s the first Sonnet model to launch with safety classifiers that prevent reasoning extraction. Sonnet 5.5 also expands preserved thinking, so Claude’s thinking cannot be decoupled from the account that created it. Most developers won’t notice a change. If you move conversations between accounts, including switching accounts mid-session in Claude Code, our docs article explains the change.
Getting started
As with Opus 5.5 and Sonnet 5, Claude Sonnet 5.5 is available with zero data retention.
Claude Sonnet 5.5 is now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can get started on the Claude Platform with claude-sonnet-5-5. If you run Sonnet with thinking off, you’ll need to switch to the new between_tools setting, which keeps up-front thinking off, before moving to Sonnet 5.5. See our migration guide for details.
Footnotes
1 Terminal-Bench 4.0 results are reported for Claude Opus 5.5 at Xhigh effort which represents the model’s highest score.
2 Sonnet 5.5 scores lower at Max effort than at Xhigh. FrontierCode evaluates whether a code change could be merged without human edits. It penalizes out-of-scope changes, even if they are high-quality or helpful. At Max effort, Sonnet 5.5 more often ran Claude Code’s code-review skill, which splits the review across many subagents, and in two cases Cognition examined, this led to a timeout or to extra edits beyond the task’s scope, and therefore to a lower score.
3 Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release deployment of Sonnet 5.5 on the Claude Platform, which we found to have a bug that could degrade responses to requests that use structured outputs. We expect the effect on Sonnet 5.5’s scores, if any, to be small and to understate its performance. That bug has since been fixed.
4 OpenAI recently fixed a bug that degraded image understanding in GPT-6 Sol. Official AA-Briefcase v1.1 and GDPval-AA v2.1 scores from Artificial Analysis, and Chartography scores from Surge AI, may not have been updated yet to reflect the latest version of the model. Artificial Analysis does not expect major impacts to AA-Briefcase v1.1 and GDPval-AA v2.1. Internal testing of Chartography suggests its score was not impacted.