本节中报告的 Claude Mythos 5 结果描述了模型在部署时未启用安全限制下的核心行为。为了反映用户接触 Claude 的不同方式,我们还报告了 Claude Fable 5 的结果,其中启用了分类器和模型切换机制,并采用了将在发布时生效的 claude.ai 系统提示词的接近最终版本。我们将定性讨论的重点放在 Mythos 5 上,因为它代表了更 [...] 次运行。Surge 对模型在全部 100 个提示词上进行了评估。
在 GDP.pdf 测试中,Claude Fable 5 达到了 29.8% 的严格通过率,优于达到 22.5% 严格通过率的 Claude Opus 4.8。Claude Fable 5 处于 GPT-5.5(得分 24.9%)和 Gemini 3.1 Pro(得分 16.7%)之上,是目前的最先进模型。
52 Surge AI. (2026). GDP.pdf: Can $100B AI models master the documents that run the world? [...] 287
-- 287 of 319 --
由 FutureHouse 开发,用于评估 AI 在实际科学研究任务中的能力。
我们在公开集的 181 个问题中对模型进行评估,并对五次运行的分数取平均值。在所有运行中,模型均配置为启用自适应思考和最大努力模式,无论是否使用 Python 工具。在使用 Python 工具进行评估时,模型会获得一个容器——其中安装了镜像文件和标准 Python 库——以及一个图像裁剪工具。
查看原文
Results reported on Claude Mythos 5 throughout this section describe the model’s core
behavior without deployment-time safeguards. To reflect the different ways users
encounter Claude, we also report results for Claude Fable 5 with its classifiers and model
switching mechanism active, and with a near-final version of the claude.ai system prompt
that will be in place at launch. We focus our qualitative discussion on Mythos 5 as the more [...] runs. Surge evaluated the model on the full 100 prompts.
On GDP.pdf, Claude Fable 5 achieved a strict pass rate of 29.8%, improving over Claude
Opus 4.8, which achieved a strict pass rate of 22.5%. Claude Fable 5 is state-of-the-art over
GPT-5.5 and Gemini 3.1 Pro, which scored 24.9% and 16.7% respectively.
52 Surge AI. (2026). GDP.pdf: Can $100B AI models master the documents that run the world? [...] 287
-- 287 of 319 --
developed by FutureHouse, which evaluates AI capabilities for practical scientific research
tasks.
We evaluate the model on 181 questions from the public set and average scores over five
runs. The model is configured with adaptive thinking and max effort enabled in all runs,
both with and without Python tools. When evaluated with Python tools, the model is
provided with a container—with the image file and standard Python libraries installed—and
an image cropping tool.