OpenAI / 长帖 / 视频
在xhigh档上,GPT-6.2 Astra的MedChemBench得分是84.5%,GPT-5.6 Sol是58.2%
这份基准有99道题,来自117篇论文和专利,包含761个分子结构。GPT-6 Astra在结构重建上拿到95.4%,选择化合物和预测数值大约只拿到一半。官方写,这不能证明模型能发现新药或改进实验室决策。
乔木改写
同一道题要求模型把一个中心结构和两个基团拼成分子,并保住一个18个原子的环。在所有11个模型共用的最高推理档xhigh上,GPT-6 Astra的10次回答全部拼对;GPT-5.6 Sol每次都拼成15个原子的环;GPT-5.6 Luna则把分子拆成互不相连的两截。
这些错误会改变正在分析的到底是哪一种分子。OpenAI把这件事写成药用化学推理的基础:模型要先正确表示化合物,再去解释它的行为。
2026年11月20日,OpenAI发布了基准MedChemBench,用来评估模型能不能把分子结构与实验证据联系起来,解释结果、比较化合物,并预测它们的性质。官方写道,做出能可靠支持这类工作的AI,是帮助科学家开发更好药物的一步。

媒体未能载入,请通过下方原文链接查看。
药用化学里很难的一点是,改善分子的一项性质,常常会把另一项弄差。一个化合物可能对疾病相关蛋白作用很强,但吸收差,或在体内分解太快。化学家会设计、制备并测试一组相关化合物,再决定下一步试什么。官方认为,AI要帮得上忙,就需要把结构变化和实验证据连起来,并判断哪些化合物值得继续做。
MedChemBench有99道题,来自117篇已发表的论文和专利,包含761个不同的分子结构。多数题目会给模型一份研究进行到某一时刻时能看到的材料:生物靶点、研究者的目标、已经研究过的化合物,以及选出的实验结果。
评分分四块。分子结构:根据提供的图纸和化学基团,拼出正确的原子连接和指定的立体化学。科学推理:解释要有题目给出的证据支持,比如为什么选中某一个化合物。排序和选择:按题目要求给化合物排序,并做出符合要求的选择。性质预测:与来源材料里报告的数值和类别一致,包括测得的和计算得到的分子性质。
在xhigh推理档上,GPT-6.2 Astra拿到了可用分数的84.5%,GPT-6 Astra是83.3%,GPT-6.1 Sol是82.0%,GPT-5.6 Sol是58.2%。官方写,GPT-6 Astra最强的是重建分子结构和解释实验证据;选择化合物和预测数值,对这个模型更难。
其中一道题来自PRMT5抑制剂研究,PRMT5是癌症药物发现里被研究过的一种蛋白。模型会拿到分子结构和更早的实验结果,要预测三个相关化合物会怎样:它们在细胞里阻断PRMT5有多强,穿过细胞膜屏障有多容易,以及在实验室检测里肝酶分解它们有多快。后来的实验没有给模型看。结果显示出权衡:化合物23在细胞里用更低浓度就能阻断PRMT5,化合物24更容易穿过细胞屏障,也在肝酶检测里分解得更慢。
官方说,把这些性质放在一起预测,是在看模型能不能把结构跟化合物行为的几个方面同时连起来。对靶点活性更强,并不一定伴随更好的通透性或更慢的分解。基准会检查分子结构有没有重建对,以及预测和实测有多接近。预测按与实测的接近程度给部分分数。题目里已经给出的测量值,也必须被正确复述。
文中的Figure 1用化合物17的几组回答,展示了正确重建、环变小,以及分子被拆开。官方在统计结构部分时写,Astra在xhigh档拿到了可用分数的95.4%。结合前后文,这里的Astra指GPT-6 Astra,因为同一段里整份基准的xhigh分数是83.3%。博客页上的这张图,以及另外两张性能图,这次没能从页面里取出可用图片地址,读者需要看原文。页面里还有演示视频,同样未能归档。
他们评估了11个GPT模型,并在不同推理力度下测试。在所有模型共用的xhigh档上,整份基准里Astra拿到83.3%,GPT-6.1 Sol是82.0%,GPT-6.1 Terra是75.1%,GPT-5.6 Sol是58.2%。在max档上,GPT-6 Astra拿到83.8%。官方写,更高的推理档会提高分数,但模型之间仍有很大差距。Figure 2画的是分数和输出token的关系:每个点是一个模型在一种推理档上的结果,共51种组合,覆盖从GPT-5.6起的9个模型;分数是99道题的平均,误差条是95%不确定区间;token包含推理过程和最终答案。
Figure 3按四个能力分开画:重建分子、解释发现、选择化合物、预测测量值。官方写,Astra最清楚的优势是重建分子结构和解释科学证据。在预印本里选出的推理例子中,Astra正确找出了结构变化,并权衡它们对蛋白结合、吸收和代谢的可能影响。这次没有找到该预印本的公开链接,这些例子以官方转述为准。
官方接着写,Astra在xhigh档的结构分数是95.4%,选择化合物和预测数值大约只拿到一半。这几类题的任务和计分规则不同,但结果表明,预测和化合物选择还有很大改进空间。
结尾他们写得很清楚:这些结果不能证明模型能发现新药,也不能证明它们能改进实验室决策。不能排除模型事先见过这些已发表来源。每次回答都是单次尝试,没有工具、没有反馈,也没有新的实验。基准评估的是根据历史证据做推理,不是真实的药物发现。官方认为,它有助于看清模型在哪些地方前进,哪些地方仍困难,并说会继续分享后续更新。
读这份基准,更值得带走的是分项分数和它自己写下的限制。总分到了80%以上,并不等于模型已经会帮科学家决定下一步该做哪个化合物。
原文附带媒体
来自OpenAI官方X宣布帖的附带视频,时长约13秒。博客正文里另有演示视频和Figure 1至Figure 3,这次未能从页面取出可用地址,不能用这条宣布视频代替。

媒体未能载入,请通过下方原文链接查看。
原文
Can AI help scientists identify promising drug candidates?

媒体未能载入,请通过下方原文链接查看。

媒体未能载入,请通过下方原文链接查看。
Can AI help scientists identify promising drug candidates?
November 20, 2026

Introducing MedChemBench, a benchmark that tests how well AI models read molecular structures, interpret experimental results, and choose which compounds to test next.
We're introducing MedChemBench, a benchmark that evaluates AI models on reasoning tasks drawn from medicinal chemistry research. It tests whether models can connect a molecule's structure with experimental evidence to explain results, compare compounds, and predict their properties. Building AI that can reliably support this work is one step towards helping scientists develop better medicines.
One of the hardest parts of this work is that improving one property of a molecule can make another worse. A compound might act strongly against a disease related protein but be poorly absorbed or break down too quickly in the body. Medicinal chemists work through these tradeoffs by designing, making, and testing related compounds, then deciding what to try next. For AI to help, it needs to connect changes in molecular structure with experimental evidence and reason about what compounds are worth pursuing.
MedChemBench measures these abilities through 99 problems drawn from published research. At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of the available credit, compared with 83.3% for GPT-6 Astra, 82.0% for GPT-6.1 Sol, and 58.2% for GPT-5.6 Sol. GPT-6 Astra was strongest at reconstructing molecular structures and explaining experimental evidence. Selecting compounds and predicting their numerical properties proved harder for that model.
What MedChemBench grades
MedChemBench draws on 117 papers and patents and includes 761 distinct molecular structures. For most problems, models receive the context available at a chosen point in a research program: the biological target, the researchers' goals, the compounds studied so far, and selected experimental findings.
The benchmark evaluates four components:
- Molecular structures. Correct atom connections and specified stereochemistry in molecules assembled from the drawings and chemical groups provided.
- Scientific reasoning. Explanations supported by the supplied evidence, such as the reasons for choosing one compound over another.
- Ranking and selection. Compound rankings and choices that meet the stated requirements.
- Property predictions. Agreement with the numerical values and categories reported in the source material, including measured and calculated molecular properties.
Prediction across competing properties
One MedChemBench problem draws on research into inhibitors of PRMT5, a protein studied in cancer drug discovery. The model receives molecular structures and results from earlier experiments. It must predict how three related compounds will behave: how strongly they block PRMT5 in cells, how readily they cross a cell barrier, and how quickly liver enzymes break them down in a laboratory assay. Later experiments, withheld from the model, showed a tradeoff. Compound 23 blocked PRMT5 in cells at a lower concentration, while compound 24 crossed the cell barrier more readily and broke down more slowly in the liver-enzyme assay.
Predicting these properties together tests whether a model can connect molecular structures with several aspects of compound behavior. This reflects a central challenge in medicinal chemistry: stronger activity against a target does not necessarily come with better permeability or slower breakdown. The benchmark checks whether the model reconstructs the molecular structures correctly and how closely its predictions match experimental measurements. Predictions receive partial credit based on their closeness to the measured values. Measurements already supplied in the prompt must also be reproduced correctly.
Small chemical changes can produce different property profiles. The task asks models to connect the structures with evidence about their behavior.
Testing spatial reasoning in medicinal chemistry
Before a model can reason about a compound's properties, it needs to identify the molecule correctly. Chemical drawings encode atom connections, ring structures, and stereochemistry.
MedChemBench tests whether models preserve this information when reconstructing molecules from the drawings and chemical groups provided.
One problem asks models to assemble a molecule from a central structure and two attached groups while preserving an 18-atom ring. At the highest reasoning setting shared by all eleven models (xhigh), GPT-6 Astra reconstructed this compound correctly in all ten responses. GPT-5.6 Sol and GPT-5.6 Luna failed across all ten responses.
In the illustrated answers, GPT-5.6 Sol produced a 15-atom ring, while GPT-5.6 Luna split the molecule into two disconnected fragments.
These errors change the identity of the molecule being analyzed. This example tests a foundation of medicinal chemistry reasoning: whether the model can accurately represent the compound before interpreting its behavior.
Figure 1. The same input, different molecules
These selected answers for compound 17 show a correct reconstruction, a smaller ring, and disconnected fragments.
While measuring the benchmark for structure components, Astra earned 95.4% of the available credit at xhigh. This result shows strong performance on reconstructing the molecular structures supplied in the tasks.
We evaluated eleven GPT models at different reasoning-effort settings. Across the full benchmark, Astra earned 83.3% of the available credit at the shared xhigh setting, compared with 82.0% for GPT-6.1 Sol, 75.1% for GPT-6.1 Terra, and 58.2% for GPT-5.6 Sol. At max effort, GPT-6 Astra reached 83.8%. Higher reasoning settings improved scores, while large differences between models remained. The figure below plots performance against output tokens.
Figure 2. Performance versus token usage
Each point represents a model at one reasoning setting, showing 51 combinations across nine models from GPT-5.6 onward. Scores are averaged across 99 problems; bars show 95% uncertainty intervals. Tokens include reasoning and the final answer.
The component results separate four abilities: reconstructing molecules, explaining findings, selecting compounds, and predicting measurements. This shows where models are strongest and where errors remain.
Figure 3. Component performance versus token usage
Astra's clearest strengths were in reconstructing molecular structures and explaining the scientific evidence. In selected reasoning examples from our preprint, Astra correctly identified structural changes and weighed their possible effects on protein binding, absorption, and metabolism.
Where selection and prediction remain difficult, Astra earned 95.4% of the available structure credit at xhigh, but only about half the credit for selecting compounds and predicting numerical properties. These categories use different tasks and scoring rules, but the results show substantial room to improve predictions and compound choices.
Where we go from here
These results do not establish whether models can discover new drugs or improve laboratory decisions. Prior exposure to the published sources cannot be ruled out, and each response was a single attempt without tools, feedback, or new experiments. While the benchmark evaluates reasoning from historical evidence rather than real-world drug discovery, its results help identify where models are advancing and where challenges remain. We believe MedChemBench will support continued progress and look forward to sharing updates as our models advance to serve the field.