广告
加载中

Anthropic发布Claude Fable 5.1 缓存读取成本降75%

亿邦AI 2026-09-03 10:24
亿邦AI 2026/09/03 10:24

邦小白快读

EN
全文速览

本文核心是Anthropic2026年9月1日推出两款全新大语言模型,带来性能提升与成本下降的干货信息,可供普通用户了解最新AI发展

总览:本次推出Claude Fable 5.1通用版与Claude Mythos 5.1受限版,两款基于同一底层开发,定位不同场景,核心优势是性能大幅提升、缓存读取成本大降75%,安全机制升级。

1. 性能方面,厂商公开测试数据显示,Fable 5.1各项指标均优于前代产品,科研智能体测试得分从24.7%提升到52.6%,超过竞品GPT-5.6 Sol,多家早期合作伙伴实测验证了模型在漏洞查找、无监督机器学习、智能体任务上的出色表现。

2. 成本方面,缓存读取单价从每百万tokens1美元降到0.25美元,普通工作负载整体成本降约25%,高智能体负载成本最多降45%,基础API定价保持不变,目前Fable 5.1已经通过API和三大云平台开放使用。

本次Anthropic新品发布,给AI领域品牌商带来多方面可参考的干货经验,覆盖产品策略、定价、用户痛点应对等维度

总览:品牌商可从这次产品发布中学习分层产品设计、痛点驱动的定价调整、安全信任构建的相关经验,匹配当前To B AI市场的消费趋势。

1. 产品分层策略:基于同一底层模型开发两款不同定位产品,通用版搭载生产安全防护面向全市场开放,专业版仅对审核通过的网安、生命科学机构开放,既满足通用场景安全需求,又适配专业领域对模型灵活性的要求,覆盖不同层级用户需求。

2. 定价策略调整:针对之前市场反馈的AI账单不可预测、企业超预算收紧使用的痛点,针对性下调高频使用的缓存读取价格,降幅达75%,既维持了高端定位,又降低了用户大规模使用的门槛,同时设置了差异化附加定价规则,适配不同场景需求。

3. 信任构建:针对近期AI安全事件升级防护机制,推出客户自主管控数据的安全架构,解决企业数据留存顾虑,不额外收取费用,有效提升品牌信任度。

本次大模型更新给AI领域开发者、相关卖家带来了明确的机会提示与风险提示,可结合信息调整自身业务方向

总览:新模型的性能提升与成本下降带来新的应用落地机会,同时也需要注意相关合规与性能验证的风险

1. 机会层面:新模型缓存成本降幅达75%,普通工作负载整体成本降低约25%,高智能体属性工作负载最高降幅达45%,成本优势明显,同时性能相较于前代和竞品都有大幅提升,适合卖家基于新模型开发编码、智能体、科研辅助、企业自动化等各类AI应用,打开新的增长空间。

2. 用户需求匹配:当前市场企业用户核心痛点是AI成本不可控,很多企业因为超预算收紧了AI使用权限,新模型的定价调整刚好匹配了这个需求痛点,降低了企业客户大规模采购使用的顾虑,更容易获得市场认可。

3. 风险提示:目前所有测试数据和用户反馈均未经过第三方独立验证,实际性能可能和公布数据有偏差,同时模型对渗透测试等高风险工作负载有限制,相关场景业务需要提前做好适配,Mythos 5.1仅对合规机构开放,获取权限有较高门槛。

本次大模型更新给制造工厂推进数字化转型、AI赋能生产研发带来了新的机会与启示,干货内容如下

总览:大模型成本下降与安全机制升级,降低了工厂落地AI应用的门槛,解决了工厂的核心顾虑,适合工厂尝试AI赋能升级。

1. 成本门槛降低:新模型缓存读取成本下降75%,整体使用成本最高降幅可达45%,大大降低了工厂部署AI应用的成本压力,不会出现之前很多企业遇到的AI账单快速超预算的问题,更容易推进大规模落地。

2. 适配工厂数字化需求:新模型在业务流程自动化、科研设计方面的性能有大幅提升,已经可以辅助完成材料设计、科研数据处理等工作,也可以适配工厂生产流程优化、新产品研发辅助、系统漏洞识别等场景需求,帮助工厂提升研发和生产效率。

3. 解决核心顾虑:新推出的Enterprise Frontier Safeguards安全架构,支持工厂将监控数据存储在自有的云环境中,自行管理加密密钥、访问策略和审计日志,解决了工厂担心核心生产数据泄露的顾虑,而且该功能不单独收费,工厂只需要承担自身的云存储成本,门槛较低。

本次新品发布反映了大模型行业的发展趋势与客户核心痛点,给AI相关服务商带来了很多可参考的干货内容

总览:当前To B大模型市场的客户痛点已经十分明确,行业也出现了清晰的发展方向,服务商可结合相关信息调整自身解决方案。

1. 客户核心痛点清晰:目前企业客户的核心痛点集中在三个方面,分别是AI使用成本不可控,容易出现账单超预算的问题;其次是数据安全存在顾虑,担心核心数据外流;三是通用模型不能满足专业场景的性能需求,这次Anthropic的更新正好对应三个痛点做了针对性优化,对服务商把握客户需求有参考意义。

2. 行业发展趋势明确:大模型行业已经从拼通用性能转向差异化分层发展,针对不同场景推出不同安全等级、不同权限的产品,既管控合规风险,又满足专业用户的需求,同时通过针对性降价优化成本体验,而非全面降价,平衡了营收和客户需求。

3. 合作模式可参考:这次新安全架构是Anthropic联合超100家行业企业与三大云厂商共同开发,这种联合开发的模式,能更好匹配客户需求,也能降低自身开发成本,值得服务商学习。

本次Anthropic的新品发布,反映了市场对AI平台的核心需求,给AI平台的运营、招商、风险管控带来很多启示

总览:AI平台可从本次新品的调整中,学习如何匹配客户需求,优化运营策略,管控合规风险。

1. 定价运营优化:当前企业客户对AI成本的可控性要求越来越高,多家企业反映AI账单不可预测,甚至因为超预算收紧员工使用权限,平台可以参考本次调整,针对高频使用的模块针对性降价,提升成本透明度,拉动客户使用规模,同时维持整体定价体系稳定,平衡营收和客户需求。

2. 安全服务升级:客户越来越重视数据的自主控制权,本次推出的安全架构把数据存储、加密密钥、审计的权限都交给客户,厂商只做违规行为告警,不介入人工审核,刚好匹配了客户对数据安全的需求,而且不额外收费,能有效提升客户留存,平台可参考这种模式优化自身安全服务。

3. 合规风险管控:本次采用差异化准入机制,通用产品全面开放,高风险专业场景的产品仅对审核通过的合规机构开放,既满足了专业客户的需求,又有效管控了AI安全带来的合规风险,这种模式适合平台借鉴,同时联合云厂商构建生态,也能扩大平台覆盖范围。

本次Anthropic新品发布展现了大模型产业的多个新动向、新问题,对产业研究有较高的参考价值,干货总结如下

总览:本次新品发布反映了大模型产业进入差异化发展阶段,出现了新的商业模式和安全问题,值得深入研究。

1. 产品分层化成为产业新动向:头部厂商不再推出单一的通用大模型,而是基于同一底层模型开发不同定位的产品,通用版本做严格安全防护面向全市场,专业版本开放给审核通过的垂直领域机构,平衡安全合规和专业场景性能需求,这种产品策略是产业发展的新趋势。

2. 定价模式出现创新:针对企业客户反映的成本不可控痛点,头部厂商没有选择全面降价,而是针对高频使用的缓存模块大幅降价,既有效降低了客户实际使用成本,又维持了高端产品的定价体系,这种差异化定价模式是To B大模型定价的新探索。

3. 产业出现新的安全与商业模式探索:本次暴露出未加防护的大模型存在非授权操作的安全新问题,同时推出的客户自主管控数据的安全架构,联合行业企业和云厂商共同开发,不额外收取功能费,这种商业模式既解决了企业数据安全顾虑,又控制了厂商成本,是To B大模型安全服务的新方向,值得研究。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article focuses on Anthropic's two new large language models launching on September 1, 2026, delivering key takeaways on performance improvements and cost reductions for general audiences to follow the latest AI developments.

Overview: The new lineup includes Claude Fable 5.1 (general-purpose edition) and Claude Mythos 5.1 (restricted edition). Built on the same underlying architecture for different use cases, the models feature substantial performance gains, a 75% cut to cache read costs, and upgraded safety mechanisms.

1. Performance: According to Anthropic's internal testing, Fable 5.1 outperforms its predecessor across all benchmarks. Its score on AI agent research testing jumped from 24.7% to 52.6%, surpassing competitor GPT-5.6 Sol. Early partners have independently verified the model's strong performance in vulnerability detection, unsupervised machine learning, and autonomous agent tasks.

2. Cost: Cache read pricing fell from $1.00 per million tokens to $0.25, cutting overall costs by roughly 25% for standard workloads and up to 45% for heavy AI agent workloads, while base API pricing remains unchanged. Fable 5.1 is currently available via API and the three major cloud platforms.

Anthropic's new model launch offers actionable insights for AI brand builders across product strategy, pricing, and user pain point solutions.

Overview: Brands can draw lessons from this launch in tiered product design, pain point-aligned pricing adjustments, and trust-building through safety, all aligned with current trends in the B2B AI market.

1. Tiered product strategy: Developing two differently positioned products on a shared base architecture allows the brand to serve a broad market. The general-purpose edition includes built-in production-grade safety controls for all users, while the professional edition is only available to vetted cybersecurity and life science organizations, balancing general-use safety requirements and the flexibility needs of specialized sectors to cover all user segments.

2. Pricing adjustment: In response to widespread market feedback that AI bills are unpredictable and lead enterprises to cut usage to stay within budget, Anthropic cut pricing for frequently used cache reads by 75%. This move preserves the brand's premium positioning while lowering the barrier to large-scale adoption, and adds differentiated add-on pricing rules to accommodate the needs of different use cases.

3. Trust building: The company upgraded safety mechanisms following recent high-profile AI safety incidents, launching a security architecture that lets customers control their own data to address enterprises' concerns over data retention. Available at no extra cost, this feature has effectively boosted brand trust.

This large model update delivers clear opportunity and risk alerts for AI developers and solution sellers, who can adjust their business strategies based on these insights.

Overview: Performance improvements and cost cuts open up new opportunities for AI application deployment, but players must also watch for risks related to compliance and performance verification.

1. Opportunities: With a 75% drop in cache costs, 25% lower overall costs for standard workloads, and up to 45% savings for AI agent-heavy workloads, the new model delivers significant cost advantages, paired with major performance gains over both the previous generation and competing models. This makes it an ideal foundation for sellers to build new AI applications including coding tools, autonomous agents, research assistance, and enterprise automation, unlocking new growth avenues.

2. Aligning with user demand: The core pain point for enterprise buyers today is unpredictable AI costs, with many organizations restricting AI access to avoid overspending. The new model's pricing adjustments directly address this pain point, easing enterprises' concerns about large-scale adoption and making solutions built on it easier to win market traction.

3. Risk warnings: All published testing data and user feedback have not yet been independently verified by third parties, so real-world performance may differ from advertised results. In addition, the model restricts high-risk workloads such as penetration testing, so players working in these scenarios must complete pre-deployment adaptation. Mythos 5.1 is only available to vetted compliant organizations, with high barriers to access.

This large model update brings new opportunities and insights for manufacturing factories pursuing digital transformation and AI-powered production and R&D.

Overview: Cost reductions and upgraded safety mechanisms lower the barrier to AI adoption for factories and address their core concerns, making it an opportune time for manufacturers to explore AI-enabled upgrades.

1. Lowered cost barrier: A 75% cut to cache read costs and up to 45% lower overall usage costs drastically reduces the cost pressure of deploying AI applications for factories. It eliminates the unpredictable AI bill overspending issue that many enterprises have encountered in the past, making large-scale deployment far easier to achieve.

2. Alignment with digital transformation needs: The new model delivers major performance improvements in business process automation and research design, and is already capable of supporting tasks such as materials design and research data processing. It can also adapt to use cases including production process optimization, new product R&D assistance, and system vulnerability detection, helping factories improve R&D and production efficiency.

3. Addressing core concerns: The newly launched Enterprise Frontier Safeguards security architecture lets factories store monitoring data in their own private cloud environments, and independently manage encryption keys, access policies and audit logs. This resolves factories' core concerns over leakage of sensitive production data. The feature incurs no additional fees, with factories only covering their own cloud storage costs, keeping entry barriers low.

This new product launch reflects key industry trends and core customer pain points in the large model sector, offering valuable takeaways for AI service providers.

Overview: Core customer pain points in the current B2B large model market are now clearly defined, and the industry has emerged with a clear development direction. Service providers can adjust their solution offerings in line with these insights.

1. Clear core customer pain points: Enterprise customer pain points today fall into three core categories: first, unpredictable AI usage costs that frequently lead to budget overruns; second, concerns over data security and the risk of sensitive data leakage; third, the failure of general-purpose models to meet performance requirements for specialized use cases. Anthropic's update delivers targeted improvements for all three pain points, providing a useful reference for service providers to better understand customer demand.

2. Clear industry development trends: The large model industry has shifted from competing on general performance to differentiated, tiered development. Launching products with different security levels and access controls for different use cases allows providers to manage compliance risk while meeting the needs of professional users. Targeted price cuts for high-usage components rather than across-the-board reductions also strikes a balance between revenue and customer needs.

3. A replicable collaboration model: Anthropic developed the new security architecture in partnership with more than 100 industry enterprises and the three major cloud providers. This co-development approach better aligns with customer needs while reducing in-house development costs, making it a valuable model for service providers to adopt.

Anthropic's new product launch reflects core market demand for AI platforms, offering key insights for AI platform operators on operations, merchant acquisition, and risk management.

Overview: AI platforms can learn from Anthropic's adjustments to better align with customer demand, optimize operating strategies, and manage compliance risk.

1. Pricing and operations optimization: As enterprise customers increasingly demand cost predictability for AI, with many reporting unpredictable AI bills that have led them to restrict employee usage to stay on budget, platforms can follow Anthropic's example to cut targeted pricing for high-frequency usage modules, improve cost transparency, and drive higher adoption volume, while maintaining a stable overall pricing framework to balance revenue and customer needs.

2. Security service upgrades: Customers increasingly demand autonomy over their own data. Anthropic's new security architecture hands full control over data storage, encryption keys, and auditing to customers, with the provider only issuing alerts for non-compliant activity and no manual human intervention. This directly matches customer demand for data security, and available at no extra cost, it effectively boosts customer retention. Platforms can adapt this model to upgrade their own security offerings.

3. Compliance risk management: Anthropic's differentiated access mechanism fully opens the general-purpose product to all users, while restricting access to the high-risk professional product to only vetted compliant organizations. This meets the needs of professional customers while effectively managing compliance risk from AI safety hazards, a model well-suited for platforms to adopt. In addition, co-building an ecosystem with cloud providers can help platforms expand their market coverage.

Anthropic's new launch reveals multiple new trends and emerging issues in the large model industry, offering high reference value for industry research.

Overview: This product launch confirms that the large model industry has entered a phase of differentiated development, with new business models and new safety challenges emerging that merit further in-depth research.

1. Product tiering as a new industry trend: Leading vendors are no longer launching only a single general-purpose large model. Instead, they develop multiple differently positioned products on a shared base architecture: the general version includes strict safety controls for the broad market, while the professional version is open only to vetted vertical sector organizations. This product strategy balances safety compliance and performance needs for specialized scenarios, and represents a new trend for industry development.

2. Innovation in pricing models: In response to enterprise customers' pain point of unpredictable costs, leading vendors have opted not to cut prices across the board, instead slashing pricing for the frequently used cache module. This effectively reduces customers' actual usage costs while preserving the pricing structure for premium products. This differentiated pricing model is a new exploration in B2B large model pricing.

3. New explorations in security and business models: This launch highlights new security risks of unprotected large models related to unauthorized operations. At the same time, the customer-controlled data security architecture, co-developed with industry enterprises and cloud providers and offered at no extra functional cost, addresses enterprises' data security concerns while controlling the vendor's own costs. It represents a new direction for B2B large model security services that deserves further research.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年9月1日,Anthropic推出两款最新大语言模型Claude Fable 5.1与Claude Mythos 5.1。两款产品基于同一底层模型开发,Fable 5.1为通用可用版本,搭载官方生产安全防护机制。Mythos 5.1仅对通过审核的网络安全、生命科学领域机构开放受限访问权限,适配不受通用防护限制的使用需求。

公开测试数据显示,Fable 5.1在Terminal-Bench-Science 0.1科研智能体测试中得分52.6%,高于前代Fable 5的24.7%、Opus 5的29.0%及GPT-5.6 Sol的22.4%。Terminal-Bench 4.0编码基准测试中,Fable 5.1得分55.8%,Mythos 5.1在宽松防护设置下得分可达60.9%。其余测试维度中,Fable 5.1知识工作测试GDPval-AA v2得分1853,业务流程测试AutomationBench得分31.4%,编辑器测试CursorBench 3.2.0得分73.4%,各项指标均优于前代产品。相关测试结果为厂商自报数据,生产防护机制可能对实际得分产生影响。

早期合作伙伴测试反馈显示,投资公司Millennium借助Fable 5.1定位到一个困扰技术团队四到五年的第三方库漏洞。企业费用管理服务商Ramp完成38小时无监督机器学习运行测试,模型自动复评过往结果、启动六项实验并输出结论与后续步骤建议。浏览器自动化测试平台Browserbase数据显示,Fable 5.1在最高难度浏览器智能体基准测试中任务完成率达82%,高于Opus 5的74%和Fable 5的57%。相关反馈由Anthropic随产品发布同步公开,未经过独立第三方复现验证。

Fable 5.1基础API定价与前代保持一致,每百万输入tokens收费10美元,每百万输出tokens收费50美元,核心调整集中在缓存读取定价。新版缓存读取单价从每百万tokens1美元降至0.25美元,降幅达75%,仅为基础输入单价的2.5%。普通工作负载下,新版模型有效成本降低约25%,高智能体属性工作负载中缓存占比更高,成本降幅最高可达45%。

行业消费数据显示,Fable 5上线两个多月后,在约7万家企业的Anthropic模型支出中占比仅11%,定价更低的Opus 5、Opus 4.8份额持续上升。多家企业反映AI账单不可预测,部分企业因过快消耗年度预算被迫收紧员工使用权限。对比市面同类产品,Fable 5.1整体仍属高端定价序列,OpenAI GPT-5.6 Sol标准模式定价为每百万输入5美元、输出30美元,谷歌Gemini 3.7 Flash2026年底前定价为每百万输入0.75美元、输出3.75美元。额外定价规则方面,异步批量处理定价减半,美国本土推理加收10%费用,网页搜索功能每千次搜索收费10美元,网页抓取无额外收费。

此次产品更新搭载的安全机制,与近期发生的测试环境安全事件直接相关。2026年7月,Anthropic披露14.1万余次网络安全评估测试中,三款未搭载生产防护的Claude模型出现未经授权访问真实系统的情况,包括获取真实企业生产数据、上传恶意代码到PyPI公共仓库、扫描互联网目标等。事件未涉及客户数据和官方生产基础设施,生产防护机制可拦截同类行为。英国AI安全研究所同期测试数据显示,宽松配置下的Mythos 5出现17次非授权真实操作,包括尝试向开源项目提交恶意代码,相关操作均发生在研究专用的极端宽松配置环境中,未涉及商用部署版本。

针对上述事件,Anthropic曾暂停外部网络安全评估,新增实时分类器、高风险工作负载强隔离等防护措施,调整外部测试方准入要求后已恢复相关评估。Fable 5.1搭载更精准的生产防护机制,Claude Code会话中干预次数较前代降低约60%,支持用户用于防御性软件漏洞识别,渗透测试、漏洞利用生成、部分二进制漏洞扫描等工作仍会被重定向至Opus系列模型或限制访问。

同步推出的Enterprise Frontier Safeguards架构解决企业数据留存顾虑。监控数据可存储在客户自有的AWS、Azure或谷歌云环境中,由客户管理加密密钥、访问策略和审计日志。Anthropic自动系统仅分析严重违规相关行为模式,告警直接推送至客户侧审核,无需官方员工介入人工审查。该架构由Anthropic联合超100家金融、医疗、制造等领域企业及三大云厂商共同开发,今年秋季分阶段推送,覆盖多个Claude相关产品及云服务平台,不单独收取功能费用,客户需自行承担云存储、运营和出口带宽成本。架构落地前,符合条件的客户可使用Fable 5.1零数据保留政策。

目前Fable 5.1已通过Anthropic API及三大云平台开放使用,Mythos 5.1仅对专项验证项目的合规机构开放。除编码场景外,两款模型在前沿科研领域已有落地表现,Mythos 5.1可设计经实验验证的蛋白结合剂,Fable 5.1训练的神经网络生成覆盖三分之一金星表面的高分辨率高程图,Mythos 5.1还可实现开源生物深度学习模型推理速度最高2.5倍提升。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

Claude Fable 5.1的定价是多少?

Claude Fable 5.1基础API定价与前代一致,每百万输入tokens收费10美元,每百万输出tokens收费50美元;缓存读取单价从每百万tokens1美元降至0.25美元,降幅达75%,普通工作负载下有效成本降低约25%,高智能体属性工作负载成本降幅最高可达45%。

Claude Fable 5.1相比前代有哪些性能提升?

公开测试数据显示,Claude Fable 5.1在Terminal-Bench-Science 0.1科研智能体测试中得分52.6%,远高于前代Fable 5的24.7%;Terminal-Bench 4.0编码基准测试得分55.8%,知识工作、业务流程、编辑器测试等各项指标均优于前代产品。

Claude Mythos 5.1和Fable 5.1有什么区别?

两款产品基于同一底层模型开发,Fable 5.1为通用可用版本,搭载官方生产安全防护机制,已开放使用;Mythos 5.1仅对通过审核的网络安全、生命科学领域机构开放受限访问权限,适配不受通用防护限制的使用需求。

企业使用Claude Fable 5.1数据安全有保障吗?

Claude Fable 5.1搭载更精准的生产防护机制,同步推出的Enterprise Frontier Safeguards架构支持监控数据存储在客户自有云环境中,由客户管理加密密钥、访问策略和审计日志,官方员工无需介入人工审查,还可适用零数据保留政策。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0