[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-claude-sonnet-5-5-terminal-bench":3,"topics-all":38,"news-related-090954a9-c27b-46c0-a864-a10fe10a2331":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"090954a9-c27b-46c0-a864-a10fe10a2331","70.6% 的中端模型:Sonnet 5.5 终端评测越级压过 Opus 5.5","Anthropic 发布 Claude 5.5 家族第二款 Sonnet 5.5:输出快三成以上,单任务成本最多省 30%,token 单价不变。它在 Terminal-Bench 4.0 拿 70.6%,超过旗舰 Opus 5.5,多项基准逼近 Opus,还是首个自带网络防护与防蒸馏分类器的 Sonnet。","Anthropic 在 9 月 28 日放出了 Claude 5.5 家族的第二款模型 Sonnet 5.5,距离旗舰 Opus 5.5 只隔六天。官方口径很直白:比 Sonnet 5 输出快 30% 以上、单任务成本最多省 30%,token 单价一分没降。真正值得注意的是另一件事——在 Terminal-Bench 4.0 这个智能体编码评测上,这款中端模型拿到 70.6%,超过了自家旗舰 Opus 5.5 的 66.4%([官方公告](https:\u002F\u002Fwww.anthropic.com\u002Fclaude-sonnet-5-5),[AI\u002FTLDR 独立收录](https:\u002F\u002Fai-tldr.dev\u002Fmodels\u002Fclaude-sonnet-5-5\u002F))。\n\n## 跑分:中端档咬住旗舰\n\n具体数字:Terminal-Bench 4.0 上 Sonnet 5.5 得分 70.6%,上一代 Sonnet 5 只有 10.3%,Opus 5.5 为 66.4%(Xhigh 档成绩)。CursorBench 4.0 拿 55.5%,距 Opus 5.5 的 57.8% 只差约两分;FrontierCode 1.1 在 Xhigh 档 52.1%,略低于 Opus 的 54.4%,高于 GPT-6 Sol 的 49.3%。知识工作类同样贴身:GDPval-AA 拿 1844 分,Opus 是 1846,Sonnet 5 只有 1449。图表识别 Chartography 61.6%,电脑操作 OSWorld 2.1 达 80.1%(partial 口径)。\n\n不过 Anthropic 自己也压了预期:在复杂、开放式、需要持续判断的任务上,Opus 5.5 依然明显更强,基准分数只反映能力的一个切面。一个有趣的注脚:官方称这是第一个只靠截图就通关《宝可梦 红》的 Sonnet。\n\n## 成本:单价不动,靠省 token 降价\n\n定价与 Sonnet 5 完全一致:输入每百万 2 美元、输出 10 美元、缓存读取 0.2 美元,是 Opus 5.5(输入 4 美元、输出 20 美元)的一半。降价全部来自 token 效率:官方测试里同等工作量最多省 30%;Medium 档就能以不到十分之一的单任务成本,超过 Sonnet 5 的最好 Terminal-Bench 成绩。\n\n第三方数据更直观:Balyasny 资产管理在 2441 个金融任务上,每个答案约用 12.1 万 token,Sonnet 5 要 49.7 万;Slack 的离线评测省了约 14% 输出 token;Base44 的 118 次真实应用构建平均 3.6 轮迭代完成,而 Opus 5 需要 7.7 轮。再配合 Low 到 Max 的 effort 分档,中端模型第一次把「调档」变成了比「换模型」更顺手的成本杠杆。\n\n## 安全:中端第一次带上旗舰级护栏\n\n因为网络安全能力与 Opus 5 相当,Sonnet 5.5 成为第一个发布即带网络防护与回退机制的 Sonnet——高风险网安任务会显式回落到 Sonnet 5。它也是第一个内置防蒸馏分类器、阻止推理能力被批量提取的 Sonnet,并扩展了 preserved thinking,让思考过程无法与创建它的账号解绑。在约 1850 个场景的自动化行为审计里,它是 Anthropic 全系模型里最不主动试探容器边界的那个。\n\n## 所以呢\n\n5.5 家族的分工已经写明:Opus 做架构与判断,Sonnet 做实现与日常量,Haiku 5.5 几周内补上低价位。当基准分数趋同、token 效率拉开数倍差距,「买中端还是买旗舰」正在变成「先买中端,不够再升档」。回头看一眼你的工作流:有多少任务,真的需要旗舰?","https:\u002F\u002Fwww.anthropic.com\u002Fclaude-sonnet-5-5\u002F","1001cead-0639-4c04-ab47-f19d863be5f2",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"4ca6514d-d6e5-40b0-b03b-2d95649b1f2b","en","Claude Sonnet 5.5: Mid-Tier Tops Opus on Terminal-Bench","Anthropic ships Sonnet 5.5: 30%+ faster, up to 30% cheaper per task, 70.6% Terminal-Bench 4.0 beats Opus 5.5. First Sonnet with cyber safeguards.","Anthropic shipped Claude Sonnet 5.5, the second model in the Claude 5.5 family, on September 28 — six days after the flagship Opus 5.5. The official pitch is blunt: output more than 30% faster than Sonnet 5, up to 30% cheaper per task, with per-token prices unchanged. The real headline is something else — on Terminal-Bench 4.0, an agentic coding evaluation, this mid-tier model scored 70.6%, ahead of its own flagship Opus 5.5 at 66.4% ([official announcement](https:\u002F\u002Fwww.anthropic.com\u002Fclaude-sonnet-5-5), [independent AI\u002FTLDR listing](https:\u002F\u002Fai-tldr.dev\u002Fmodels\u002Fclaude-sonnet-5-5\u002F)).\n\n## Benchmarks: mid-tier bites the flagship\n\nThe numbers: 70.6% on Terminal-Bench 4.0, versus 10.3% for Sonnet 5 and 66.4% for Opus 5.5 (Xhigh effort). On CursorBench 4.0 it scores 55.5%, within about two points of Opus 5.5's 57.8%. On FrontierCode 1.1 it reaches 52.1% at Xhigh — slightly below Opus at 54.4%, above GPT-6 Sol at 49.3%. Knowledge work is just as tight: GDPval-AA at 1844, versus Opus's 1846 and Sonnet 5's 1449. Chartography chart recognition lands at 61.6%; OSWorld 2.1 computer use at 80.1% (partial mode).\n\nAnthropic itself tempers expectations: for complex, open-ended work requiring sustained judgment, Opus 5.5 remains clearly stronger, and benchmarks capture only one facet of capability. One fun footnote: the company says it is the first Sonnet model to beat Pokémon Red working only from screenshots.\n\n## Cost: same sticker price, cheaper by token efficiency\n\nPricing is identical to Sonnet 5: $2 per million input tokens, $10 per million output, $0.20 for cache reads — half of Opus 5.5's $4 input \u002F $20 output. The discount comes entirely from token efficiency: in Anthropic's testing, the same work costs up to 30% less, and at Medium effort it beats Sonnet 5's best Terminal-Bench score for less than a tenth of the cost per task.\n\nThird-party data makes the point sharper. Balyasny Asset Management, across 2,441 finance tasks, saw roughly 121k tokens per answer where Sonnet 5 needed 497k. Slack's offline evals used about 14% fewer output tokens. Base44's 118 real app builds finished in 3.6 iterations on average, where Opus 5 took 7.7. Combined with Low-to-Max effort levels, the mid-tier model turns \"adjusting the dial\" into a cheaper lever than \"switching models.\"\n\n## Safety: the first mid-tier with flagship guardrails\n\nBecause its cybersecurity capabilities are comparable to Opus 5's, Sonnet 5.5 is the first Sonnet to launch with cyber safeguards and fallbacks — higher-risk security tasks visibly fall back to Sonnet 5. It is also the first Sonnet with safety classifiers that block reasoning extraction, and it expands preserved thinking so a model's thinking cannot be decoupled from the account that created it. In an automated behavioral audit of roughly 1,850 scenarios, it was the least likely of Anthropic's models to probe the limits of its container.\n\n## So what?\n\nThe 5.5 family's division of labor is now explicit: Opus sets architecture and judgment, Sonnet executes everyday volume, and Haiku 5.5 fills the budget tier within weeks. When benchmark scores converge and token efficiency diverges by multiples, \"flagship-first\" deserves a second look. Look at your own workload: how many tasks truly need the flagship?","claude-sonnet-5-5-terminal-bench","2026-09-28T23:05:32Z","2026-09-28T23:10:49.950228Z","2026-09-28T23:10:49.950237Z",true,"agent",742,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"6cd8c9b7-b493-4ad7-bd20-a907b9fb70f4","Claude 5.5 家族首发:Opus 5.5 缓存读取降六成,典型负载省四成","claude-opus-5-5-cost-cut-benchmarks","2026-09-22T21:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"63b28ef5-3ffa-4897-88d2-5dcd7fb678b5","Claude Fable 5.1 发布:缓存读取降价 75%,Agent 科研基准翻倍","claude-fable-5-1-mythos-release","2026-09-02T13:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a5c82cbd-def9-4d33-99b3-cba1d7ab4abe","Haiku 5.5 发布:输入一毛钱,跑分最高翻4.6倍","claude-haiku-5-5-launch","2026-10-07T21:09:42+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"96ebc302-b442-44fe-a23d-d08befbc1246","Saturn 报告:18 款主流 AI 答理财错 57%","saturn-ai-finance-benchmark-57-percent","2026-09-26T05:04:42+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5930aa08-8a81-4c92-9ec0-79c3f4de39e2","Anthropic 推出 Claude Opus 5:性能逼近 Fable 5,API 价格不变继续啃企业市场","claude-opus-5-launch","2026-07-25T02:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"20904504-4450-4f2e-a612-49b191403a86","Claude Opus 5 thinking 默认开启：OSWorld 70.57% 的慢即是快","claude-opus-5-thinking-default","2026-07-25T00:00:00+00:00"]