Anthropic 发布 Claude Fable 5.1 与 Mythos 5.1，编码更强且成本最多降 45%

人工分析（Artificial Analysis）曾协助Anthropic进行预发布测试，但对节省成本的说法表示质疑：https://www.linkedin.com/posts/artificial-analysis_claude-fable-51-tops-the-artificial-analysis-activity-7500649293469335552-6g83?utm_source=share&utm_medium=member_desktop&rcm=ACoAABfX0nABz6sCWbPldiV_9liETVfz5fRLAD0。虽然缓存读取减少在代理工作负载中每项任务节省约1.40美元，但在最大努力下，Fable 5.1 每项任务的实际成本比 Fable 5 高出20%，因为它使用了大约1.7倍的输出令牌。在最大努力下，Fable 5.1 每个智能指数任务花费3.76美元，而 Opus 5 则为2.34美元，得分仅低三分。在额外高努力下，Fable 5.1 得分65，每项任务成本为2.72美元，缩小了差距，但仍比 Opus 5 贵。

Anthropic 推出 Claude Fable 5.1 和 Mythos 5.1，这些是迄今为止其最强大的 AI 模型。除了在代理编码和文本质量上的提升外，公司还将成本削减了最高45%。

与前代模型一样，这两款5.1模型共享相同的基础模型，但在安全防护方面有所不同。Fable 5.1 面向广大用户，而 Mythos 5.1 仅限于网络安全和生命科学的特殊访问项目。

Fable 5.1 对于典型工作负载的成本比 Fable 5 低约25%，对于需要长时间自主运行、调用多个工具的高代理任务，节省幅度可达约45%。Anthropic 之所以能实现这一点，是因为将缓存读取费用从每百万令牌1美元降至0.25美元。其他所有 API 价格保持不变，即每百万输入令牌10美元、每百万输出令牌50美元。相比之下，Opus 5 的运行成本只有一半，即每百万输入令牌5美元、每百万输出令牌25美元。Ad

关于《Fable 5》的最大抱怨是高成本，这也可能导致该模型在企业客户中采用率低：https://the-decoder.com/fable-5s-slow-adoption-suggests-corporate-willingness-to-pay-for-frontier-ai-has-hit-a-ceiling/。自从 Opus 5 在七月底发布以来，价格压力一直在增加：https://the-decoder.com/anthropics-claude-opus-5-costs-well-below-fable-5-while-matching-or-beating-it-across-most-benchmarks/，其在大多数基准测试中已经与 Fable 5 持平或超过，价格却更低。像早期的 Claude 模型一样，新的版本配备了一个控制计算使用的努力等级系统。在低或中等努力等级下，Fable 5.1 应该可以在更低成本下达到 Fable 5 的效果。

Fable 5.1 在代理基准测试中获得了重大提升。在 Terminal-Bench-Science 0.1 中，该模型得分为 52.6%，是 Fable 5 的 24.7% 的两倍多，也远高于 GPT-5.6 Sol 的 22.4%。在 Terminal-Bench 4.0 的代理编程测试中，Fable 5.1 得分为 55.8%，而 Mythos 5.1 达到 60.9%，相比之下 Fable 5 为 42.0%，GPT-5.6 Sol 为 37.3%。这些提升是否能在现实世界中同等规模地转化为有效应用，将在未来几周内明朗。广告

Anthropic 研究员 Felix Rieseberg：https://x.com/felixrieseberg/status/2094849659059703818 表示，Fable 5.1 还改进了其写作风格。早期模型在聊天中过于依赖要点和加粗文本，而 Fable 5.1 在遵循风格指引的同时有所收敛，使整体表达更自然。

针对网络安全、生物学和医学问题的安全过滤器不如早期版本那么严格。在 5.0 版本中，这些过滤器过于敏感，即使是善意的请求也会触发，导致虚假正例频发，非常令人沮丧。广告

在 5.1 版本中，网络安全过滤器产生的虚假正例减少了 60%，而针对生物学相关查询，对于关于基础生物学和医学的无害问题，过滤器触发的频率减少了 85%。广告

寓言 5.1 现在可以首次识别软件漏洞，但不能开发利用程序。渗透测试和利用生成仍然会通过 Opus 模型处理。Mythos 也是 Claude Security 的一部分，用于防御目的：https://the-decoder.de/anthropic-macht-sein-staerkstes-modell-claude-mythos-5-erstmals-fuer-die-cyber-verteidigung-verfuegbar/。

Claude Fable 5.1 现已可在所有平台立即使用，包括 AWS、Google Cloud 和 Microsoft Azure。开发者可以通过 API 使用 claude-fable-5-1 访问它。对于企业客户，Anthropic 正在推出企业前沿防护（Enterprise Frontier Safeguards, EFS），该方案将客户数据仅存储在客户自己的云基础设施上。

Claude Mythos 5.1 当前仅限于美国机构，通过两个项目提供：用于防御性安全工作的网络验证计划（Cyber Verification Program）和与美国政府合作开发的生命科学验证计划（Life Sciences Verification Program）。Anthropic 计划将访问权限扩展到国际合作伙伴。

Anthropic 还在打击蒸馏攻击，即通过成千上万个虚假账户系统地提取模型能力。新的 API 账户在多轮对话中不能再编辑 Claude 以往的上下文，同时保留思考记录，据 Anthropic 称，这封堵了已知的蒸馏技术。

保持对 AI 的关注。清晰、有用，无废话。

关注 The Decoder 获取 AI 新闻、背景故事和专家分析。

解码器：https://the-decoder.com/

Artificial Analysis, which helped Anthropic with pre-release testing, disputes the savings claim：https://www.linkedin.com/posts/artificial-analysis_claude-fable-51-tops-the-artificial-analysis-activity-7500649293469335552-6g83?utm_source=share&utm_medium=member_desktop&rcm=ACoAABfX0nABz6sCWbPldiV_9liETVfz5fRLAD0. While the cache read cut saves about $1.40 per task on agentic workloads, Fable 5.1 at max effort actually costs 20 percent more per task than Fable 5 because it uses roughly 1.7 times as many output tokens. At max effort, Fable 5.1 runs $3.76 per Intelligence Index task compared to Opus 5 at $2.34, which scores only three points lower. At extra-high effort, Fable 5.1 scores 65 at $2.72 per task, closing the gap but still costing more than Opus 5.

Anthropic launches Claude Fable 5.1 and Mythos 5.1, its most capable AI models yet. Along with gains in agentic coding and text quality, the company cuts costs by up to 45 percent.

Like their predecessors, both 5.1 models share the same base model but differ in safety guardrails. Fable 5.1 is broadly available, while Mythos 5.1 is restricted to special access programs for cybersecurity and life sciences.

Fable 5.1 costs about 25 percent less than Fable 5 for typical workloads, with savings climbing to roughly 45 percent for heavily agentic tasks involving long, autonomous runs with many tool calls. Anthropic made that possible by slashing cache reads from $1 to $0.25 per million tokens. All other API prices remain unchanged at $10 per million input tokens and $50 per million output tokens. For comparison, Opus 5 runs at half that price with $5 input and $25 output per million tokens. Ad

High cost was the biggest complaint about Fable 5, and it likely contributed to the model seeing low adoption among enterprise customers：https://the-decoder.com/fable-5s-slow-adoption-suggests-corporate-willingness-to-pay-for-frontier-ai-has-hit-a-ceiling/. The price pressure has been building since Opus 5 launched：https://the-decoder.com/anthropics-claude-opus-5-costs-well-below-fable-5-while-matching-or-beating-it-across-most-benchmarks/ in late July, already matching or beating Fable 5 on most benchmarks at a lower price. Like earlier Claude models, the new versions come with an effort-level system that controls compute usage. At low or medium effort, Fable 5.1 should match Fable 5's results at lower cost.

Fable 5.1 posts major gains on agentic benchmarks. On Terminal-Bench-Science 0.1, the model hits 52.6 percent, more than double Fable 5's 24.7 percent and far ahead of GPT-5.6 Sol at 22.4 percent. On Terminal-Bench 4.0 for agentic coding, Fable 5.1 scores 55.8 percent while Mythos 5.1 reaches 60.9 percent, compared to 42.0 percent for Fable 5 and 37.3 percent for GPT-5.6 Sol. Whether these gains translate to real-world use at the same scale will become clear over the coming weeks. Ad

Anthropic researcher Felix Rieseberg：https://x.com/felixrieseberg/status/2094849659059703818 says Fable 5.1 also improves its writing style. Earlier models leaned too heavily on bullet points and bold text in chat, and Fable 5.1 dials that back while following style instructions more closely and sounding more natural overall.

The safety filters for cybersecurity, biology, and medical questions are less aggressive than in earlier versions. In the 5.0 models, they were so sensitive that they triggered on well-intentioned requests too, producing false positives at a frustrating rate. Ad

The cybersecurity filters in 5.1 generate 60 percent fewer false positives, and for biology-related queries, the filters fire 85 percent less often on harmless questions about basic biology and medicine. Ad

Fable 5.1 can now identify software vulnerabilities for the first time, though not develop exploits. Penetration testing and exploit generation still get routed to the Opus models. Mythos is also part of Claude Security for defensive purposes：https://the-decoder.de/anthropic-macht-sein-staerkstes-modell-claude-mythos-5-erstmals-fuer-die-cyber-verteidigung-verfuegbar/.

Claude Fable 5.1 is available immediately on all platforms, including AWS, Google Cloud, and Microsoft Azure. Developers can access it via the API using claude-fable-5-1 . For enterprise customers, Anthropic is rolling out Enterprise Frontier Safeguards (EFS), which store customer data solely on the customer's own cloud infrastructure.

Claude Mythos 5.1 is currently limited to US organizations through two programs: the Cyber Verification Program for defensive security work and the Life Sciences Verification Program, developed with the US government. Anthropic plans to expand access to international partners.

Anthropic is also cracking down on distillation attacks, where a model's capabilities are systematically extracted through thousands of fake accounts. New API accounts can no longer edit Claude's prior context in multi-turn conversations while keeping the thinking transcript, which according to Anthropic closes a documented distillation technique.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/