[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-claude-fable-5-1-mythos-release":3,"topics-all":38,"news-related-63b28ef5-3ffa-4897-88d2-5dcd7fb678b5":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},"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","Anthropic 于 9 月 1 日发布 Claude Fable 5.1 与受限版 Mythos 5.1。缓存读取价降至 0.25 美元\u002F百万 token,典型负载成本约降 25%;Terminal-Bench-Science 得分 52.6% 较前代翻倍,并展示蛋白质设计与金星测绘等科研能力。","9 月 1 日,Anthropic 上线了 Claude Fable 5.1 和 Claude Mythos 5.1。这两个名字背后是同一个底层模型:Fable 5.1 面向所有用户开放,Mythos 5.1 只对通过审核的网络安全和生命科学机构开放,安全护栏配置更宽松。发布时机耐人寻味——就在一周前,Ramp 的企业支出数据还在显示,售价更高的 Fable 5 上市两个多月只占到企业 Anthropic 支出的约 11%,客户更愿意用便宜型号。而 5.1 这次的重点,恰恰是降价。\n\n## 性能:科研型 Agent 基准翻倍\n\nAnthropic 公布的对比数据里,最扎眼的是 Terminal-Bench-Science 0.1(智能体科研基准):Fable 5.1 拿到 52.6%,前代 Fable 5 为 24.7%,Opus 5 为 29.0%,GPT-5.6 Sol 为 22.4%——相对前代是翻倍以上的提升。其他基准:\n\n- **Terminal-Bench 4.0**(智能体编码):Fable 5.1 得 55.8%,护栏更少的 Mythos 5.1 冲到 60.9%,而 Fable 5 仅 42.0%\n- **CursorBench 3.2.0**:Fable 5.1 得 73.4%,超过 Fable 5(70.5%)、Opus 5(70.0%)和 GPT-5.6 Sol(67.2%)\n- **Humanity's Last Exam**(无工具):60.9%,高于前代的 57.8%\n- **OSWorld 2.0**(电脑操作,strict 口径):41.7%,Opus 5 为 39.6%\n\n值得注意的是 effort 分级:Fable 5.1 在 Claude Code 默认高努力档,在 Claude.ai 和 Claude Cowork 默认中档。官方称低\u002F中档即可用低得多的成本达到或超过 Fable 5 的效果。\n\n## 定价:砍的就是 Fable 5 被吐槽的那块\n\n输入输出价格不变(输入 10 美元\u002F百万 token、输出 50 美元),但缓存读取直降 75%,降到 0.25 美元\u002F百万 token。按 Anthropic 的测算:典型负载总成本约降 25%,缓存读取占大头的高度 Agent 化工作负载最高省约 45%。对照此前\"企业不买旗舰账单\"的数据,这一刀砍在痛点上——Cognition 在官方公告中直言,靠新的缓存读取定价,Fable 级模型对他们此前一直跑在 Opus 上的工作负载终于\"经济可行\",并在发布当天就把 Devin 里的 Opus 5 流量迁移了过来。\n\n## 科研展示:从蛋白质设计到金星地图\n\n这次发布最具想象力的部分是科研成果展示:\n\n- **蛋白质设计**:Mythos 5.1 借助开源蛋白质设计与折叠工具产出的 binder,在三个靶点上结合亲和力比 Adaptyv Bio 蛋白质设计竞赛的最佳提交高 10 倍;12 个靶点的命中率接近 50%(行业典型为 10-15%),并经两家外部机构实验验证。\n- **金星地形图**:Fable 5.1 用 NASA Magellan 任务 30 多年前的雷达数据训练神经网络,绘制出金星约三分之一表面的新高程图,分辨率从过去的 10-20 公里提升到 2-3 公里,高度精度提高 25%,已按 CC 许可公开发布。\n- **GPU 内核优化**:Mythos 5.1 为 7 个开源深度学习模型编写定制 GPU 内核并缓存中间结果,推理加速最高 2.5 倍,基因组级分析的 GPU 成本估算下降 30-60%,官方称这些优化计划开源。\n\n还有个具体案例:对冲基金 Millennium 一个约百万次运行才出现一次的崩溃,自家工程师和其他模型四五年都没查明,Fable 5.1 通过反汇编外部供应商库并对照 core dump 定位了根因。\n\n## 企业隐私与护栏:数据不出客户的云\n\n配套发布的企业前沿护栏(Enterprise Frontier Safeguards,EFS)允许客户把数据存在自己控制的云基础设施上,人工审查默认由客户自己完成——等效零数据保留的同时保留滥用检测,今秋起分阶段上线。网络安全方向的护栏干预减少约 60%,并首次允许用 Fable 5.1 做软件漏洞发现(仍禁止开发利用代码)。此外,针对工业化蒸馏攻击的反制措施也进一步收紧。\n\n## 所以呢\n\n表面上这是一次小版本升级,实际上是 Anthropic 对\"旗舰太贵、企业不买账\"的直接回应:性能往上走一步,成本往下砍一刀,再用科研案例证明贵有贵的道理。真正的看点在于,当缓存读取价格降到原来的四分之一,长时程 Agent 任务的经济账会被整个重算——你上个月评估过\"跑不起\"的那类工作流,现在值得重新核一遍成本。\n\n参考:Anthropic 官方公告(https:\u002F\u002Fwww.anthropic.com\u002Fclaude-fable-and-mythos-5-1)","https:\u002F\u002Fwww.anthropic.com\u002Fclaude-fable-and-mythos-5-1","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},"b29f09a1-11e7-4a1d-9a48-2bc146239b17","en","Claude Fable 5.1: Cache Reads Cut 75%, Agentic Benchmarks Double","Anthropic ships Claude Fable 5.1: cache reads cut 75% to $0.25\u002FM tokens, Terminal-Bench-Science doubles to 52.6%, with protein design and Venus mapping demos.","On September 1, Anthropic launched Claude Fable 5.1 alongside Claude Mythos 5.1. Both names point to the same underlying model: Fable 5.1 is generally available, while Mythos 5.1 is restricted to vetted cybersecurity and life-sciences organizations through trusted-access programs, with more permissive safeguards. The timing is telling — just a week earlier, Ramp's spending data showed Fable 5 holding only about 11% of enterprise Anthropic spend more than two months after launch, with customers favoring cheaper models. The theme of this release is, squarely, price.\n\n## Performance: agentic science benchmarks double\n\nThe standout number in Anthropic's comparison table is Terminal-Bench-Science 0.1, an agentic scientific-research benchmark: Fable 5.1 scores 52.6%, versus 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol — more than double its predecessor. Other results:\n\n- **Terminal-Bench 4.0** (agentic coding): Fable 5.1 hits 55.8%, Mythos 5.1 (with fewer safeguard interventions) reaches 60.9%, while Fable 5 sits at 42.0%\n- **CursorBench 3.2.0**: Fable 5.1 scores 73.4%, ahead of Fable 5 (70.5%), Opus 5 (70.0%), and GPT-5.6 Sol (67.2%)\n- **Humanity's Last Exam** (no tools): 60.9%, up from 57.8%\n- **OSWorld 2.0** (computer use, strict): 41.7%, versus 39.6% for Opus 5\n\nThe release also introduces per-message effort levels: Fable 5.1 defaults to High effort in Claude Code and Medium on Claude.ai and Claude Cowork. Anthropic says Low or Medium effort matches or beats Fable 5's results at much lower cost.\n\n## Pricing: cutting exactly where Fable 5 was criticized\n\nInput and output prices are unchanged ($10 per million input tokens, $50 per million output), but cache reads drop 75% to $0.25 per million tokens. By Anthropic's own math, typical workloads cost about 25% less overall, and highly agentic workloads — where cache reads dominate — save up to roughly 45%. Given the earlier \"enterprises won't pay for the flagship\" data, this cut lands on the pain point: Cognition stated in the announcement that with the new cache-read pricing, a Fable-class model finally became economical for workloads it had kept on Opus, and it moved its Devin traffic from Opus 5 on launch day.\n\n## Science demos: from protein design to Venus\n\nThe most imaginative part of the release is the scientific research showcase:\n\n- **Protein design**: given open-source protein design and folding tools, Mythos 5.1 produced binders with 10x higher binding affinity than the best designs submitted to Adaptyv Bio's protein design competitions on three targets; its hit rate reached nearly 50% across 12 targets (10–15% is typical), validated experimentally by two external organizations.\n- **Venus elevation map**: Fable 5.1 trained a neural network on radar images from NASA's Magellan mission — collected more than 30 years ago — to produce a new elevation map covering roughly a third of Venus, improving resolution from 10–20 km down to 2–3 km with heights up to 25% more accurate, released under a Creative Commons license.\n- **GPU kernel optimization**: Mythos 5.1 wrote custom GPU kernels for seven open-source deep learning models, speeding up inference by up to 2.5x and cutting estimated GPU costs for genome-wide analyses by 30–60%; Anthropic says it plans to open-source these optimizations.\n\nThere's also a concrete anecdote: a roughly one-in-a-million crash at hedge fund Millennium that its engineers and every other model had failed to explain for several years was root-caused by Fable 5.1, which disassembled an external vendor library and matched it against the core dump.\n\n## Enterprise privacy and safeguards: data stays in the customer's cloud\n\nThe accompanying Enterprise Frontier Safeguards (EFS) let customers store data on cloud infrastructure they control, with human review done by the customer by default — zero-data-retention-equivalent privacy while retaining misuse detection, rolling out in phases starting this fall. Cybersecurity safeguard interventions drop by about 60%, and Fable 5.1 can now be used to discover software vulnerabilities (developing exploits remains prohibited). Anti-distillation mechanisms were also tightened further.\n\n## So what\n\nOn the surface this is a point release. In practice it is Anthropic's direct answer to \"the flagship is too expensive and enterprises won't buy it\": push performance up a step, cut costs with a cleaver, and use science demos to justify the premium. The real story is that when cache-read prices fall to a quarter of what they were, the economics of long-horizon agentic work get recalculated from scratch — workflows you evaluated last month and dismissed as unaffordable are worth re-costing today.\n\nReference: Anthropic's official announcement (https:\u002F\u002Fwww.anthropic.com\u002Fclaude-fable-and-mythos-5-1)","claude-fable-5-1-mythos-release","2026-09-02T13:20:00Z","2026-09-02T13:09:36.862816Z","2026-09-02T13:09:36.862826Z",true,"agent",237,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"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",{"id":70,"title":71,"news_slug":72,"published_at":73},"428c2822-a53a-4075-80ec-2960e03ea062","Anthropic Claude Sonnet 5：中端档拉到 Opus 4.8 水位","claude-sonnet-5-main-model","2026-07-01T02:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"aa29e468-6de8-417b-99f3-efe024635249","Claude Fable 5 落地：Mythos 级能力首次走向大众，\"自动路由\"改写安全发布规则","claude-fable-5-mythos-auto-routing","2026-06-09T03:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d4c2c83a-f1ea-4a1e-848e-5f1169e9d42b","微软MAI-Thinking-1：清洁训练的35B MoE推理模型，对位Claude Opus 4.6与Sonnet 4.6","mai-thinking-1-microsoft-35b-moe-clean","2026-06-02T06:14:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5f699f9d-2750-42f8-956f-918507cf51b2","Claude Opus 4.5 发布：Anthropic 夺回编程能力榜首位置","claude-opus-4-5-coding-crown-reclaim","2026-05-29T11:05:00+00:00"]