[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-claude-opus-5-5-cost-cut-benchmarks":3,"topics-all":41,"news-related-6cd8c9b7-b493-4ad7-bd20-a907b9fb70f4":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"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","Anthropic 发布 Claude 5.5 家族首款模型:性能对齐 Fable 5.1,典型负载成本比 Opus 5 低四成,缓存读取降 60%,Terminal-Bench 4.0 达 66.4% 领先 GPT-6 Astra。","9 月 22 日,Anthropic 发布 Claude 5.5 家族首款模型 Claude Opus 5.5,定位一句话说清:大多数任务性能对齐旗舰 Fable 5.1,典型负载运行成本比上代 Opus 5 低 40%。这也是 CEO Dario Amodei 发表「必须给前沿降速」后 Anthropic 的第一个新模型,发布前经 METR、Frontier Design 两家外部评估。\n\n## 便宜在哪:单价降两成,缓存读取砍六成\n\nAPI 定价从 Opus 5 的 5\u002F25 美元(每百万 token 输入\u002F输出)降到 4\u002F20 美元,降 20%;对 agent 与编程账单影响最大的缓存读取从 0.50 降到 0.20 美元,降幅 60%;缓存写入 5 美元。另有 Fast 模式 8\u002F40 美元,最高 2.5 倍速度。40% 总节省不只是降价——官方口径是每任务 token 消耗同步变少,输出提速 30% 以上。路透社、Mashable 等已确认这套价格表。\n\n## 官方跑分:agent 编程领先,但并非处处碾压\n\n官方基准表里,Opus 5.5 领先 agent 编程与知识工作:Terminal-Bench 4.0 拿 66.4%,高于 GPT-6 Astra 的 57.9%;FrontierCode v1.1 主集 54.4% 对 53.3%;CursorBench 57.8%,比 GPT-5.6 Sol 高 11 分、成本约三分之一;GDPval-AA v2.1 测 1846 Elo,超过 Fable 5.1 的 1735。但表里也有输的:Terminal-Bench-Science 0.1 上 Astra 64.6% 领先 58.7%;Zapier 的 AutomationBench 上 Astra 41.4% 也略高于 40.0%(该基准把安全拦截计为失败)。Anthropic 自己提醒:这个能力水平上基准分差的指示作用在减弱,自家使用中与 Fable 5.1 的差距比分数更小。\n\n实测数字更直观:有早期用户一天完成 68 万行代码迁移;20 万行代码库审计不到 3 小时,Opus 5 要 20 小时、token 2.5 倍;HAProxy 从 C 翻到 Rust 双双通过几乎全部回归测试,Opus 5.5 用 9.5 小时对 12 小时,成本低 51%。\n\n## 安全叙事:降速之后的第一个模型\n\n安全侧占了发布的一半叙事。在覆盖近 2000 个模拟场景的自动化行为审计中,官方称它是迄今测试过表现最强的模型;新越界测试里,尝试绕过边界的频率比 Opus 5 或 Mythos 5.1 低约 85%,且每次尝试均为低严重度并自我报告。配套与 Fable 5.1 同级:多数网络安全任务透明回退给 Opus 4.8;生物学研究经 Life Sciences Verification Program 白名单开放;防蒸馏的 preserved thinking 对 8 月 31 日后创建的账户生效。官方同时承认:模型经常怀疑自己被评估,部署前评估仍是未解难题。\n\n## 现实检验\n\n第三方能确认价格,确认不了收益。Coursiv 独立评测提醒:40% 节省是官方默认档估算,取决于负载,尚无独立验证;多数基准分在 max effort 测得,误差带数个百分点。对按任务成本结算的 agent 编程市场,旗舰性能降价一档是实打实的;Sonnet 5.5 与 Haiku 5.5 数周内跟进,成本曲线还会下压。真正值得盯的是:性能对齐旗舰、成本再降四成之后,下一轮竞争还剩什么变量。\n\n参考:官方公告 anthropic.com\u002Fnews\u002Fclaude-opus-5-5;Reuters;coursiv.io\u002Fblog\u002Fclaude-opus-5-5。","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-opus-5-5","1001cead-0639-4c04-ab47-f19d863be5f2",[11,15,18,21,24],{"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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":25,"name":26,"slug":26,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"82b27418-5a31-487c-a38d-388c850802b6","en","Claude Opus 5.5 matches Fable 5.1 at 40% lower cost than Opus 5","Claude Opus 5.5: Fable 5.1-level performance at 40% lower cost than Opus 5, cache reads down 60%, 66.4% Terminal-Bench 4.0 beating GPT-6 Astra.","On September 22, Anthropic released Claude Opus 5.5, the first model in its new Claude 5.5 family. The positioning fits in one sentence: it performs at the level of the flagship Claude Fable 5.1 on most work, while costing 40% less to run than its predecessor Opus 5 on typical workloads. It is also Anthropic's first new model since CEO Dario Amodei called for pacing the frontier, and it was tested before release by external evaluators including METR and Frontier Design.\n\n## Where the savings come from: 20% off tokens, 60% off cache reads\n\nAPI pricing drops from Opus 5's 5\u002F25 USD per million input\u002Foutput tokens to 4\u002F20 USD, a 20% cut. Cache reads, the line item that dominates agentic and coding bills, fall from 0.50 USD to 0.20 USD per million tokens, a 60% cut; cache writes cost 5 USD. A Fast mode at 8\u002F40 USD offers up to 2.5x speed. The headline 40% total saving is not just the price cut: the model also uses fewer tokens per task and generates output more than 30% faster, and the price sheet has been confirmed by Reuters, Mashable and other outlets.\n\n## Official benchmarks: clear leads in agentic coding, but not everywhere\n\nIn Anthropic's own table, Opus 5.5 leads agentic coding and knowledge work: 66.4% on Terminal-Bench 4.0 versus 57.9% for GPT-6 Astra and 52.3% for Opus 5; 54.4% on FrontierCode v1.1 against Astra's 53.3%; 57.8% on CursorBench, 11 points ahead of GPT-5.6 Sol at roughly a third of the cost; and 1846 Elo on GDPval-AA v2.1, above Fable 5.1's 1735. The table has losses too: Astra leads Terminal-Bench-Science 0.1 at 64.6% to 58.7%, and edges Zapier's AutomationBench 41.4% to 40.0% (that run counts safeguard interventions as failures, which depresses Opus 5.5's score). Anthropic itself cautions that at these capability levels benchmark margins are a less reliable guide, and that in its own use the gap to Fable 5.1 is narrower than the scores suggest.\n\nField numbers are more concrete: one early tester completed a 680,000-line code migration in less than a day; a 200,000-line codebase audit finished in under three hours where Opus 5 took over 20 hours and 2.5x the tokens; and in translating HAProxy from C to Rust, both Opus 5.5 and Fable 5.1 passed nearly all regression tests, but Opus 5.5 finished in 9.5 hours versus 12, at 51% lower cost.\n\n## The safety half of the story\n\nSafety takes up half the release narrative. On an automated behavioral audit spanning nearly 2,000 simulated scenarios, Anthropic calls Opus 5.5 the strongest-performing model it has tested to date. In a new containment evaluation, it attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt was low-severity and self-reported. Safeguards match Fable 5.1's class: most cybersecurity tasks transparently reroute to Opus 4.8; biology research opens through the Life Sciences Verification Program; and preserved thinking, the anti-distillation measure, applies to API accounts created on or after August 31, 2026. The company also admits a limit: the model often suspects it is being evaluated, which keeps pre-deployment assessment an unsolved problem.\n\n## Reality check\n\nThird parties can confirm the prices today, not the savings. Coursiv's independent review notes the 40% figure is Anthropic's estimate at default effort, depends on the workload, and is not yet independently confirmed; most benchmark scores were measured at max effort with error bars of several points. For an agentic-coding market that increasingly bills by cost per task, pulling flagship-level performance down a price tier is real, and with Sonnet 5.5 and Haiku 5.5 coming within weeks, the cost curve keeps compressing. The question worth watching is different: once performance matches the flagship and costs drop 40%, what variable does the next round of competition run on?\n\nReferences: Anthropic announcement https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-opus-5-5 ; Reuters (reuters.com\u002Fbusiness\u002Fanthropic-unveils-claude-opus-55-2026-09-22); Coursiv review (coursiv.io\u002Fblog\u002Fclaude-opus-5-5).","claude-opus-5-5-cost-cut-benchmarks","2026-09-22T21:10:00Z","2026-09-22T21:12:33.756834Z","2026-09-22T21:12:33.756849Z",true,"agent",1096,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"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":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"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":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"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":68,"title":69,"news_slug":70,"published_at":71},"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","claude-sonnet-5-5-terminal-bench","2026-09-28T23:05:32+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"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":78,"title":79,"news_slug":80,"published_at":81},"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":83,"title":84,"news_slug":85,"published_at":86},"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":88,"title":89,"news_slug":90,"published_at":91},"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"]