OpenAI发布湿实验室基准BioBasicsBench，GPT-6 Astra得66.7%
人工智能能不能读懂真实的实验室证据？OpenAI 与酶设计公司 Ligo 合作推出了一项新基准 BioBasicsBench，用 100 项任务考查 AI 智能体解读原始湿实验室证据的能力：凝胶图像、菌落平板、显微镜照片、测序读数与仪器导出数据。
在最高推理设置下，GPT-6 Astra 获得了 66.7% 的量表分，GPT-5.6 Sol 为 43.8%，评分基于 18,500 次尝试。OpenAI 表示，这项基准意在检验 AI 能否在真实科研场景中处理杂乱的实验数据，而非只在演示环境中作答。
不过，发布方同时列出了这项测试的局限：分数是 OpenAI 自己的评分，66.7% 是部分量表分、不是通过率；发布时没有公布人类基线，所测模型均为 OpenAI 自家模型、未测试其他公司的模型；数据集与评分标准在发布时未公开链接。
OpenAI Releases Wet-Lab Benchmark BioBasicsBench; GPT-6 Astra Scores 66.7%
Can AI read real laboratory evidence? OpenAI, working with enzyme-design company Ligo, has introduced BioBasicsBench, a benchmark of 100 tasks that tests whether AI agents can interpret raw wet-lab evidence: gel images, colony plates, microscopy, sequencing reads and instrument exports.
At the highest reasoning setting, GPT-6 Astra earned 66.7% of the available rubric credit, against 43.8% for GPT-5.6 Sol, across 18,500 graded attempts. OpenAI says the benchmark is meant to test whether AI can handle messy experimental data in real research settings, not just answer in a demo.
The release also carries limits: the scores are OpenAI's own; 66.7% is partial rubric credit, not a pass rate; no human baseline was published, all seven models tested were OpenAI's own and no other company's models were tested, and the dataset and rubrics were not publicly linked at launch.
来源：OpenAI
发布：GSM·美国（发布链接）
图片来源：图：实验室研究人员进行样本操作（美国疾控中心图片，公共领域）。A researcher at work in a laboratory. (CDC, public domain)
GSM Global · 本篇英文版 English edition of this story
