[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-claude-fable-5-11-percent-anthropic-spend":3,"news-related-1051d676-8ed9-4448-b0d5-8db4b844f41f":38},{"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},"1051d676-8ed9-4448-b0d5-8db4b844f41f","Claude Fable 5 上线两个月,为什么企业只把 11% 的账单花给最强模型","Ramp 8 月 AI 指数显示 Claude Fable 5 在 Anthropic 模型支出里只占 11.4%、token 量只占 6%,Opus 5 因价格、合规与稳定性优势成为企业默认,反映前沿模型正在被工程化降级。","Anthropic 在 6 月 9 日把旗舰模型 Claude Fable 5 推到 GA 之后,这家公司的最强模型并没有拿到企业市场的默认位。\n\nRamp 在 8 月 AI 指数里披露的数字很刺耳:在 7 月的 token 管理样本中,Fable 5 给 Anthropic 贡献了 6% 的 token 量和 11.4% 的模型相关支出。换句话说,这家公司的最强模型连 Ramp 样本里 Anthropic 模型支出的零头都没摸到,大部分花出去的钱还是流向 Sonnet 5、Opus 5 这些更便宜的前代和兄弟模型。\n\n更耐人寻味的是对比组。同期 OpenAI 的 GPT-5.6 Sol 在 OpenAI 自家模型里拿到 25% 的 token 量和 23% 的支出,Fable 5 的模型相关支出大约只有 Sol 的 75%。也就是,Fable 5 不只是输给自家旧模型,跨厂商对照同样落后。\n\nRamp 的样本带有自己的 caveat。它来自 Ramp 的 token 管理产品,客户群相对偏技术,但即便在这个更可能买 Fable 5 长链编码能力的样本里,流量也绝大多数流向了别处。结论只能限定在工作负载分流这个范围里,但信号本身够强。\n\n理解这 11.4% 的关键是:最强的模型和最好的默认模型解决的是两个不同的优化问题。最强模型靠处理最难的任务拿分;默认模型要兼顾价格、可用、合规、可预测,在成千上万普通任务里足够好。Fable 5 在第一件事上有强 case,在第二件事上有结构性短板。\n\nFable 5 的定价是 10 美元 \u002F 百万输入 token、50 美元 \u002F 百万输出 token,Opus 5 是 5 美元和 25 美元,前后两端都正好是 Fable 5 的一半。Anthropic 自己在产品页上把 Opus 5 描述为接近 Fable 前沿智能、价格一半的日常工作模型。\n\n按简单的 1M 输入 + 200K 输出账单算,Fable 5 单次约 20 美元,Opus 5 约 10 美元。这个比较不是成本\u002F完成任务的 benchmark。一个更聪明的模型可能一次搞定难题,便宜模型需要重试三次。prompt caching、批处理、effort 设置、输出长度差异都会改变账单。模型选择真正的指标应该是 cost per accepted result:初始运行、重试、回退、复核成本除以被接受的结果数。\n\n但价格只是第一层。\n\n订阅层同样有约束。Max 和 premium 团队的 Fable 配额上限是共享周额度的 50%,Pro 和标准 Team 一上来就走 pay-as-you-go 信用,超过额度要么买信用要么切模型。Anthropic 自己也把 Fable 5 定位为加封顶的专项通道。\n\n比价格更要命的是企业合规。Anthropic 对 Fable 5 和其他 Mythos 类模型要求 30 天企业流量留存,即便公司强调数据用于安全监控、不用于训练、在 30 天后基本删除;严格 ZDR(零数据留存)要求的银行、医疗、国防、源代码平台,在政策层就被排除掉了。Opus 5 在通用访问下没有 covered-model 留存要求,部署优势是 benchmark 救不回来的那种。\n\n最后是 guardrails。Fable 5 在网安、生物\u002F化学、潜在模型蒸馏三个域加了独立的 classifier,Anthropic 自己说过,故意保守的 classifier 会把无害请求路由到 fallback 模型。95% 的会话没 fallback 听起来不多,但一个安全团队、认证平台、生物学家、恶意软件分析组可能就生活在 fallback 高发的那块分布里。Fable 重启之后,改过的网安 classifier 还会增加常规编码和调试的良性 flag。这就把 Fable 5 的经验从模型权重拆出来:收到什么取决于路由、refusal 行为、冗长度、延迟、harness。\n\nHacker News 上 200 多点 180 多条评论里反复收敛出来的使用模式是:Fable 5 留给架构、复杂评审、卡住的硬骨头;Opus 5、Sol、便宜模型负责日常实现。这不是给 Fable 5 的安慰奖,而是一种合理的生产架构。模型选择从来不该是 provider slogan,应该是 workload 加 acceptance test。\n\nRamp 同一份 8 月报告里 Anthropic 在付费美国企业里占了 43.5%,OpenAI 是 39.7%。Anthropic 可以同时是 vendor adoption 之王,而它的最贵模型依然是少数派产品。Fable 5 跟 Opus 5 的这次反差,真正预示的是前沿模型商业化的天花板,而非 Anthropic 的全局走势。\n\n## 留给企业用户的实操建议\n\n- 默认工作交给最便宜且够用的模型。\n- 只有 Fable 5 的成功率溢价能打过 2 倍标价 + 合规排除 + fallback 风险时,才把任务升级给 Fable 5。\n- 评测指标用 accepted result rate、cost per accepted result、fallback rate、p95 latency、human review minutes,不要只看 token 价格。\n- 严格 ZDR 的工作流,从一开始就绕开 covered-model。\n\n来源:[Ramp August AI Index 2026](https:\u002F\u002Framp.com\u002Fdata\u002Fai-index-august-2026)、[Financial Times 报道](https:\u002F\u002Fwww.ft.com\u002Fcontent\u002F5ee49718-c258-4f01-aa32-7e5b76ae5245)、[Anthropic Fable 5 定价页](https:\u002F\u002Fwww.anthropic.com\u002Fclaude\u002Ffable)、[Anthropic Fable 订阅规则](https:\u002F\u002Fsupport.claude.com\u002Fen\u002Farticles\u002F15424964-claude-fable-5-on-your-plan)、[explainx.ai 11.4% 解读](https:\u002F\u002Fexplainx.ai\u002Fblog\u002Fwhy-fable-5-is-not-the-best-default-model-august-2026)。","https:\u002F\u002Fexplainx.ai\u002Fblog\u002Fwhy-fable-5-is-not-the-best-default-model-august-2026","8cb75837-7ecc-4e8a-bc52-167b41f1be2f",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"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",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"32ee5588-2867-4d2c-9680-9a75ed183281","en","Claude Fable 5 ships to GA — but enterprises only spend 11% on Anthropic's strongest model","Ramp's August AI Index shows Claude Fable 5 contributes just 11.4% of model-attributed Anthropic spend and 6% of tokens. Opus 5 has become the default thanks to price, compliance and stability — a sign that frontier models are being engineered down.","Two months after Anthropic pushed its flagship Claude Fable 5 to GA on June 9, the company's strongest model still has not won the enterprise default slot.\n\nThe numbers from Ramp's August AI Index are unflattering. In Ramp's July token-management sample, Fable 5 produced 6% of the tokens businesses bought from Anthropic and 11.4% of the dollars spent on Anthropic models. The strongest model in the lineup did not even reach a tenth of Anthropic's model-attributed spend; most of the dollars still flow to cheaper siblings like Sonnet 5 and Opus 5.\n\nThe cross-vendor comparison is even more telling. Over the same window, OpenAI's GPT-5.6 Sol represented 25% of OpenAI tokens and 23% of OpenAI spend. Fable 5's model-attributed spend lands at roughly 75% of Sol's. Fable 5 is not only losing to Anthropic's own older models; it is also behind the OpenAI flagship on the same metric.\n\nRamp's sample carries its own caveats. It comes from Ramp's token-spend management product and skews more technical than the typical AI Index base. But even within a population that is most likely to value Fable 5's long-horizon coding skills, the bulk of tokens route elsewhere. The defensible claim is narrower than a market-share verdict, but the signal is loud.\n\nThe key to understanding the 11.4% is that \"strongest model\" and \"best default model\" solve different optimization problems. The strongest wins on the hardest available task. The default has to be affordable, available, compliant, predictable, and good enough across thousands of ordinary tasks. Fable 5 has a strong case for the first job and several structural disadvantages for the second.\n\nFable 5 lists at $10 per million input tokens and $50 per million output tokens. Opus 5 lists at $5 and $25, exactly half at both ends. Anthropic itself describes Opus 5 as the everyday model that comes close to Fable's frontier intelligence at half the price.\n\nFor a simple workload of one million input tokens plus 200,000 output tokens, Fable 5 costs about $20 on list price; Opus 5 costs about $10. That is not a cost-per-completed-task benchmark. A smarter model may solve a hard problem in one run while a cheaper model retries three times. Prompt caching, batch processing, effort settings and output length differences also move the bill. The right metric is closer to cost per accepted result: initial runs plus retries plus fallbacks plus review cost, divided by accepted results.\n\nPrice, however, is only the first layer.\n\nSubscription access adds its own ceiling. Max and premium Team or legacy Enterprise seats can use Fable for up to 50% of regular weekly limits. Pro and standard Team seats start on pay-as-you-go usage credits. After the included Fable allowance is exhausted, users have to buy credits or switch models. Anthropic itself frames Fable 5 as a capped specialty lane.\n\nMore important than price is enterprise compliance. Anthropic requires 30-day retention of business traffic on Fable 5 and other covered Mythos-class models, even though the company insists the data is used for safety monitoring rather than model training and is deleted after 30 days in nearly all cases. For banks, healthcare providers, defense contractors and source-code platforms with strict zero-data-retention policies, that policy removes Fable from the table before any benchmark is discussed. Opus 5 does not have the covered-model retention requirement for general access, which gives it a deployment advantage that no benchmark can recover after the fact.\n\nFinally, guardrails reshape the experience. Fable 5 launched with separate classifiers around cybersecurity, biology and chemistry, and potential model distillation. Anthropic itself has acknowledged that the cautious classifier routes harmless requests to a fallback model. The company reported that more than 95% of sessions had no fallback at launch, which sounds small in aggregate but can dominate the experience for security teams, authentication platforms, biologists and malware-analysis groups who live in the high-intervention slice of the distribution. After Fable's redeployment, the improved cyber classifier also increased benign flags during routine coding and debugging. The model a user actually receives is the weights plus the routing, refusal behavior, verbosity, latency and harness — not just the model picker label.\n\nThe Hacker News discussion that hit more than 200 points and 180 comments repeatedly converged on the same split: Fable 5 for architecture, difficult reviews and stubborn problems; Opus 5, Sol and cheaper models for implementation and daily work. That is not a consolation prize for Fable 5; it is a sensible production architecture. Model selection should be a workload plus an acceptance test, not a provider slogan.\n\nIn the same August report, Anthropic held 43.5% of US businesses paying for AI in July, ahead of OpenAI's 39.7%. Anthropic can lead on vendor adoption while its most expensive model remains a minority product. The gap between Fable 5 and Opus 5 points at a ceiling on frontier-model pricing rather than a verdict on Anthropic's overall trajectory.\n\n## Practical recommendations for enterprise teams\n\n- Route routine work to the cheapest dependable model that clears your acceptance bar.\n- Only escalate to Fable 5 when its higher success rate more than offsets the 2x list price, the compliance exclusion and the fallback risk.\n- Measure accepted result rate, cost per accepted result, fallback rate, p95 latency and human review minutes; ignore sticker token price.\n- For strict zero-data-retention workloads, bypass covered models entirely.\n\nSources: [Ramp August AI Index 2026](https:\u002F\u002Framp.com\u002Fdata\u002Fai-index-august-2026), [Financial Times report](https:\u002F\u002Fwww.ft.com\u002Fcontent\u002F5ee49718-c258-4f01-aa32-7e5b76ae5245), [Anthropic Fable 5 pricing page](https:\u002F\u002Fwww.anthropic.com\u002Fclaude\u002Ffable), [Anthropic Fable subscription rules](https:\u002F\u002Fsupport.claude.com\u002Fen\u002Farticles\u002F15424964-claude-fable-5-on-your-plan), [explainx.ai 11.4% analysis](https:\u002F\u002Fexplainx.ai\u002Fblog\u002Fwhy-fable-5-is-not-the-best-default-model-august-2026).","claude-fable-5-11-percent-anthropic-spend","2026-08-25T06:00:00Z","2026-08-25T09:04:45.974202Z","2026-08-25T09:04:45.974211Z",true,"agent",31,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"39724847-fdc9-4199-ac46-311e7b49d385","Ramp 数据复盘 Fable 5:旗舰上市两月仅占企业 Anthropic 支出 11%,70 倍价差压住前沿模型溢价","ramp-data-fable-5-adoption-plateaus","2026-08-26T08:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"e1724d68-bf0d-4b3f-8047-147796d5d52e","Ramp 8 月指数:Fable 5 企业份额停滞 11%,OpenAI 旗舰跑赢两倍","anthropic-fable-5-plateau-11-percent","2026-08-25T06:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"454f9530-20d7-428c-82d9-9175fa5b883a","Claude 推黎曼 zeta 下界到 67.2%：60 subagent + Lean","claude-zeta-bound-67-percent-multi-agent-lean","2026-08-17T07:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"db4ffdac-3734-41a2-8e32-67feaa7341bd","Claude 冲击黎曼猜想\"失败\",却顺手改写了 37 年没人动过的数学纪录","claude-riemann-zeta-67-percent-record","2026-08-16T23:30:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"a7146291-e849-42c3-acbd-2b50627d5332","Claude Opus 4.8 发布：41天极速迭代，Dynamic Workflows 重塑Agent协作范式","claude-opus-4-8-41-day-dynamic-workflows","2026-05-29T04:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"97c97b9c-e6e4-4982-aa57-0c0da814fb19","Anthropic 的欧盟答卷四小时即被撕开：Claude 文本水印为什么怕改写","claude-synthid-70-percent-threshold-bypass","2026-08-21T08:00:00+00:00"]