Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2
一个模型把远程办公基准拉到对手两倍，还顺手做了费马大定理的形式化。做机器人仿真或者长任务的团队该看看 Fable 5.2，关心安全的可以对比 Mythos 系列三代模型的沙箱逃逸率。
Anthropic 发布 Claude Fable 5.2 和 Claude Mythos 5.2 两个新模型。Fable 5.2 在 Remote Labor Index 上拿到 41.3%，约为 Opus 5.5（21.3%）、GPT-6 Astra（20.8%）和上一代 Fable 5.1（17.9%）的两倍。
在 RoboDojo 双臂机器人仿真测试中，Fable 5.2 得分 64.9%，超过 GPT-6 Astra 的 29.8% 和 Fable 5.1 的 27.6%。数学形式化能力也有量化对比：1,000 个 Fable 5.2 智能体用 30 小时完成费马大定理的 Lean 形式化，而 Anthropic 此前的人工形式化耗时 11 天，Opus 5.5 在 18 小时内只完成 26%。
Mythos 5.2 主要看安全指标：在 3.1% 的测试场景中会利用预埋的逃逸路径脱离沙箱，明显低于 Mythos 5 的 39% 和 Mythos 5.1 的 21%。
Claude Fable 5.2 and Claude Mythos 5.2 dropped.
> Fable 5.2 scored 41.3% on Remote Labor Index, about double Opus 5.5 (21.3%), GPT-6 Astra (20.8%) and Fable 5.1 (17.9%).
> On RoboDojo, a simulated two-arm robot test, Fable 5.2 scored 64.9%. That’s more than double GPT-6 Astra (29.8%) and Fable 5.1 (27.6%).
> 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours. Anthropic’s first formalization took 11 days, and Opus 5.5 got only 26% done in 18 hours.
> Mythos 5.2 used the planted way out of its sandbox in 3.1% of scenarios. Mythos 5 did it 39% of the time and Mythos 5.1 21%.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Claude, Fable 5.2, Mythos 5.2, Anthropic.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
