Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2，多项能力超越 Opus 5.5
费马大定理证明被 1000 个智能体形式化这事挺离谱的，而且 Fable 5.2 用一半成本追平了 Opus 5.5 的成绩。做编程和科研的可以看看这组基准数据。
Anthropic 宣布推出 Claude Fable 5.2 和 Mythos 5.2 两款模型，官方称二者在编程、推理和视觉等几乎所有评测能力上超过 Opus 5.5。输出速度比 Fable 5.1 和 Mythos 5.1 快最多 30%，表达更清晰。
基准成绩上，Fable 5.2 在 FrontierCode 2.0 的 High effort 档位以约一半的单任务成本追平 Opus 5.5 的 Max effort 成绩；在 Xhigh 档位则拿到 Anthropic 测试的已发布模型中的最高分。1,000 个 Fable 5.2 智能体在 30 小时内将费马大定理的既有证明形式化并转为机器可检验的形式，同等设置下 Opus 5.5 在 18 小时内完成 26%。Mythos 5.2 准确建模了 34 个被测衣壳（用于递送基因疗法的蛋白质外壳）中的 28 个。
两款模型在多数对齐评测上与近期 Claude 模型持平或更优，且发布前经过外部评估机构测试。Medium effort 现为 Claude 应用与 Claude Code 中的默认档位。
LETS GO: Anthropic just announced Claude Fable 5.2 and Mythos 5.2. And its insane!
According to the company, both outperform Opus 5.5 in almost every capability it evaluated, including coding, reasoning and vision.
TL;DR:
- Up to 30% faster output than Fable 5.1 and Mythos 5.1, with clearer communication.
- On FrontierCode 2.0, Fable 5.2 at High effort matched Opus 5.5 at Max effort for roughly half the cost per task. At Xhigh, it had the highest score among the released models Anthropic tested.
- 1,000 Fable 5.2 agents formalized the existing proof of Fermat’s Last Theorem in 30 hours, translating it into a form a computer can check. Opus 5.5 completed 26% in 18 hours with the same setup.
- Mythos 5.2 accurately modeled 28 of 34 tested capsids, the protein shells used to deliver gene therapies.
- Both models performed as well as or better than recent Claude models on most alignment measures and were tested by external evaluators before release.
Medium effort is the default in Claude apps and Claude Code.
本条情报归属 AITOP「模型」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Claude Fable 5.2、Mythos 5.2、Anthropic、推理模型，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「kimmonismus」（@kimmonismus），经 AITOP 多维评分管线结构化提炼。
