Anthropic 发布了 Claude Fable 5.1 和 Claude Mythos 5.1，距 Fable 5 系列于 2026 年 6 月推出已过去三个月。这两个模型是同一底层模型，只是安全防护层不同。Fable 5.1 已作为 claude-fable-5-1 全面开放；Mythos 5.1 仍仅限经过审核的组织使用。两者均支持 1M token 上下文窗口和 128K 最大输出 token，且自适应思考始终开启。最亮眼的性能数字是在 Terminal-Bench-Science 0.1 上取得 52.6% 的成绩，而 Fable 5 为 24.7%，Opus 5 为 29.0%。最亮眼的商业数字是缓存读取价格下调 75%，从每百万 token 1.00 美元降至 0.25 美元，Anthropic 测算这使典型工作负载成本降低约 25%，智能体类工作负载最高可降低 45%。基础输入和输出定价保持不变，分别为每百万 token 10 美元和 50 美元。
可以部署吗？
可以，Claude Fable 5.1 已作为 claude-fable-5-1 在 Claude API、Amazon Bedrock、AWS 上的 Claude Platform、Google Cloud 和 Microsoft Foundry 全面开放。Claude Mythos 5.1 则不行：它仅限 Project Glasswing 内经过审核的美国组织使用。
基准测试数据
在智能体科学研究基准 Terminal-Bench-Science 0.1 上，Fable 5.1 得分 52.6%，而 Opus 5 为 29.0%，Fable 5 为 24.7%，GPT-5.6 Sol 为 22.4%。Anthropic 报告每个模型的标准误差为 3.5 到 4.5 个百分点，因此应关注差距幅度，而非排名本身。
在 Terminal-Bench 4.0 上，Fable 5.1 达到 55.8%，Mythos 5.1 达到 60.9%。两个相同模型之间的差距就是安全防护干预的成本，这是一次异常坦诚的披露。其他方面：CursorBench 3.2.0 为 73.4%，Humanity's Last Exam 在无工具条件下为 60.9%、有工具条件下为 65.0%，AutomationBench 为 31.4%，OSWorld 2.0 严格模式为 41.7%，GDPval-AA v2 为 1853。
成本削减从何而来
基础输入和输出定价保持不变。缓存读取价格下降 75%，从每百万 token 1.00 美元降至 0.25 美元，即基础输入价格的 0.025 倍，而其他所有 Claude 模型均为 0.1 倍。Anthropic 测算典型工作负载成本降低约 25%，上下文密集的智能体类工作负载最高可降低约 45%。批处理价格为每百万 token 5 美元和 25 美元。
团队将遇到的三个破坏性变更
- 强制工具调用已取消：将 tool_choice 设为 any 或 tool 会返回 400 错误。请改用 auto 配合严格工具调用或结构化输出。
- 思考块与模型绑定：Fable 5.1 能读取早期模型的思考内容，但没有任何早期模型能读取它自己的思考。路由器和回退设置在降级切换时会丢失推理能力。
- 编辑早期对话轮次会使思考块失效：在对话中途注入和删除每轮提醒，或重建系统提示词或工具数组，现在都会报错。此检查仅针对 2026 年 8 月 31 日当天或之后创建的账户生效。修复方案是使用按轮次作用域的系统消息和服务器端上下文编辑。
增量变更：按消息设置推理努力程度、按轮次作用域的系统消息，以及 thinking.display: "updates" 均处于测试阶段，需通过请求头启用。内容溯源为强制要求，所有输出均带有统计文本水印，文件附带 C2PA 凭证。
Anthropic 也记录了真实的性能回退。并行工具调用的稳定性有所下降，因此智能体循环可能每轮只发起一次调用，而 Fable 5 可以批量发起多次。模型叙述性内容减少，在低推理努力程度下更常直接凭记忆作答，并且倾向于整文件重写而非针对性编辑。
安全防护与科学研究
网络安全防护现在允许漏洞发现，但不允许漏洞利用开发，这使得 Claude Code 中的干预次数每会话减少约 60%。生物安全防护在良性请求上的触发频率降低了 85%。渗透测试、漏洞利用生成和基于二进制的漏洞扫描仍会重定向到 Opus。
在研究方面，Mythos 5.1 设计的蛋白质结合剂在 12 个靶点上的命中率约为 50%，而常规水平为 10% 至 15%；Fable 5.1 构建了分辨率为 2 至 3 公里的金星高程图；自定义 GPU 内核将七个开源基因组学模型的速度提升了最高 2.5 倍。
核心要点
- Fable 5.1 和 Mythos 5.1 是同一模型外加两层安全防护；只有 Fable 全面开放可用。
- 在 Terminal-Bench-Science 0.1 上取得 52.6% 的成绩，约为 Fable 5 的 24.7% 的两倍。
- 缓存读取费用下降 75%，至每百万 token 0.25 美元；基础费率保持不变。
- 三项破坏性 API 变更会影响所有编辑对话历史的智能体。
- 数据保留期为 30 天；零数据保留需要明确授权。
Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, three months after the Fable 5 line shipped in June 2026. The two are the same underlying model behind different safeguard layers. Fable 5.1 is generally available as claude-fable-5-1; Mythos 5.1 stays restricted to vetted organizations. Both carry a 1M token context window and 128K max output tokens, with adaptive thinking always on. The headline capability number is 52.6% on Terminal-Bench-Science 0.1, against 24.7% for Fable 5 and 29.0% for Opus 5. The headline commercial number is a 75% cut to cache reads, from $1.00 to $0.25 per million tokens, which Anthropic measures as roughly 25% lower cost on typical workloads and up to 45% on agentic ones. Base input and output pricing is unchanged at $10 and $50 per million.
Is it deployable?
Yes, Claude Fable 5.1 is generally available as claude-fable-5-1 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Claude Mythos 5.1 is not: it is restricted to vetted US organizations inside Project Glasswing.
The benchmark numbers
On Terminal-Bench-Science 0.1, an agentic scientific research benchmark, Fable 5.1 scores 52.6% against 29.0% for Opus 5, 24.7% for Fable 5, and 22.4% for GPT-5.6 Sol. Anthropic reports a standard error of 3.5 to 4.5 points per model, so treat the margin, not the ranking, with care.
On Terminal-Bench 4.0, Fable 5.1 reaches 55.8% and Mythos 5.1 reaches 60.9%. The gap between two identical models is the cost of safeguard interventions, which is an unusually honest disclosure. Elsewhere: CursorBench 3.2.0 at 73.4%, Humanity’s Last Exam at 60.9% without tools and 65.0% with tools, AutomationBench at 31.4%, OSWorld 2.0 at 41.7% strict, and GDPval-AA v2 at 1853.
Where the cost cut comes from
Base input and output pricing is unchanged. Cache reads drop 75%, from $1.00 to $0.25 per million tokens, which is 0.025 times base input against 0.1 on every other Claude model. Anthropic measures roughly 25% lower cost on typical workloads and up to about 45% on context-heavy agentic ones. Batch processing is $5 and $25 per million tokens.
Three breaking changes teams will hit
- Forced tool use is gone:
tool_choiceset toanyortoolreturns a 400. Useautowith strict tool use or structured outputs instead. - Thinking blocks are model-bound: Fable 5.1 reads earlier models’ thinking, but no earlier model reads its own. Router and fallback setups lose reasoning when they switch down.
- Editing earlier turns invalidates thinking blocks: Injecting and deleting per-turn reminders, or rebuilding the
systemortoolsarray mid-conversation, now errors. The check is enforced for accounts created on or after August 31, 2026. The fixes are turn-scoped system messages and server-side context editing.
Additive changes: per-message effort, turn-scoped system messages, and thinking.display: "updates" are all in beta behind headers. Content provenance is not optional, with a statistical text watermark on all output and C2PA credentials on files.
Anthropic also documents real regressions. Parallel tool calling is more variable, so agent loops may issue one call per turn where Fable 5 batched several. The model narrates less, answers from memory more often at low effort, and prefers whole-file rewrites over targeted edits.
Safeguards and science
Cyber safeguards now permit vulnerability discovery but not exploit development, cutting interventions in Claude Code by roughly 60% per session. Biology safeguards fire 85% less often on benign requests. Penetration testing, exploit generation, and binary-based vulnerability scanning still redirect to Opus.
On research, Mythos 5.1 designed protein binders with roughly 50% hit rate across 12 targets against a 10 to 15% norm, Fable 5.1 built a Venus elevation map at 2 to 3 km resolution, and custom GPU kernels sped up seven open-source genomics models by up to 2.5x.
Key Takeaways
- Fable 5.1 and Mythos 5.1 are one model behind two safeguard layers; only Fable is generally available.
- 52.6% on Terminal-Bench-Science 0.1, roughly double Fable 5’s 24.7%.
- Cache reads fall 75% to $0.25 per million; base rates are untouched.
- Three breaking API changes hit any agent that edits conversation history.
- 30-day retention applies; zero data retention needs express authorization.