OpenAI's MedChemBench: AI reads drug papers well, but this skill still trips it up
OpenAI's MedChemBench tests AI on real drug-discovery problems from research papers. GPT-6.2 Astra scores 84.5%, yet struggles at picking drug candidates and predicting properties — an open gap for chemistry-minded developers.
On 20 November 2026, OpenAI published a test where AI models had to read molecule drawings from real research papers. Then they had to decide which drug candidate a lab should test next. The open question is which part of a drug chemist's work AI still cannot do well, and that gap is where a student can build skills.
OpenAI's MedChemBench tests four things a medicinal chemist does, and AI is strong at rebuilding structures but weak at selecting compounds and predicting properties. GPT-6.2 Astra tops it at 84.5% of available credit, according to OpenAI's announcement. Models read molecular structures well but earn only about half credit when picking compounds and predicting numerical properties. OpenAI's summary gives results for only three of the four abilities, and scientific reasoning is not singled out. That split shows where a science-curious developer can add value.
MedChemBench: what OpenAI released and when
OpenAI introduced MedChemBench on 20 November 2026 as a public benchmark of how well AI handles early drug-discovery work, and its announcement page explains the full design.
A benchmark is a fixed set of problems that every model answers, so their scores can be compared fairly. A medicinal chemist is a scientist who designs and improves molecules that could become medicines.
The test covers 761 molecular structures across its 99 problems, all taken from 117 published papers and patents (OpenAI). That means the questions come from real lab work, not textbook exercises.
Four abilities the benchmark grades
Each problem checks one of four skills: rebuilding a molecular structure, scientific reasoning, ranking or selecting compounds, and predicting numerical properties, as listed on OpenAI's page.
These names sound abstract. Here is what each one looks like as an ordinary task at a chemist's desk.
| Ability | What the model must do | How it fares |
|---|---|---|
| Rebuilding structures | Papers often draw one core shape, then list the changed side pieces in a table. The model must rebuild each full molecule, like assembling a car from a base model plus an options list. | Strong (OpenAI) |
| Scientific reasoning | Read lab results and explain why one small change made a compound work better, like working out why one tweak to a recipe fixed the taste. | Not singled out in OpenAI's summary |
| Ranking and selecting | Look at several candidates and choose which to make next, like picking which few dishes to cook when ingredients are limited. | Weak, about half credit (OpenAI) |
| Predicting properties | Estimate a measured number for a molecule that the paper does not show, like guessing a dish's calorie count from its recipe. | Weak, about half credit (OpenAI) |
One useful detail for students: the problems come from published papers. You can open the same research a model was tested on and try the task yourself.
Scores: GPT-6.2 Astra leads at 84.5%
GPT-6.2 Astra scored 84.5% at the xhigh reasoning setting, just ahead of GPT-6 Astra and GPT-6.1 Sol, and far above GPT-5.6 Sol, according to OpenAI's results.
The "xhigh" setting is the option where a model is allowed to think the longest before it answers. "Share of available credit" means the points a model earned out of all the points the problems offered.
The big jump is between generations, not between the three newest models. Those sit within a few points of each other, so the exact version name matters less than the model family. If you want to know what the top model costs to use, I covered that in GPT-6.2 Astra and Sol pricing for students.
This test covers one narrow job. For which models developers call most in everyday work, check OpenRouter's usage rankings.
Reading structures is strong, predicting properties is weak
GPT-6 Astra earned 95.4% on rebuilding structures at xhigh, but only about half credit on selecting compounds and predicting numerical properties, a split that Interesting Engineering's independent report also confirms.
A molecular property is a measurable trait of a compound. A simple example is how well it dissolves in water. A drug that will not dissolve has trouble getting from a pill into your blood, so chemists care about this number early.
My reading of why this part is harder: rebuilding a structure means copying what is already on the page. Predicting a property means estimating a number the page never shows. That needs a real model of how atoms behave, not careful reading.
What the benchmark does not prove
OpenAI says the results do not show that models can discover new drugs or improve lab decisions, a limit that Superpower Daily's coverage puts front and center.
The test conditions were narrow. As I read OpenAI's setup, each model made a single attempt, with no tools and no feedback. A working chemist checks databases, runs software and learns from failed experiments.
There is a second problem. The papers are public, so the models may have read them during training. Some correct answers could come from memory rather than reasoning, and OpenAI cannot rule that out. The claim also does not rest on OpenAI's word alone: Interesting Engineering flags the same limit.
What it shows
- Models can read chemistry papers and rebuild molecules accurately.
- Newer model families improve clearly on older ones.
- Where the weak spots are, measured on real research.
What it cannot show
- That AI can find a new drug.
- How models perform with tools, retries or lab feedback.
- Whether answers came from reasoning or from memory of the papers.
Drug-discovery ML as a career path in India and the US
Drug-discovery and cheminformatics machine learning sits where chemistry and programming overlap, and the open weak spot is a clear problem to study.
Cheminformatics means using software to store, search and analyse chemical data. The benchmark tells you which part of that work is already handled well. Reading structures from papers looks close to solved. Picking candidates and predicting properties are still open problems, and working on them means learning both the chemistry and the machine learning.
This path suits a student who likes both chemistry and code. Treat it as a project path, not a quick route to a first developer job: roles in drug-discovery ML often ask for graduate study, and this post has no hiring figures to offer. A project in this area still shows you can handle chemical data.
Say you are in your final year, choosing a project. A chatbot clone teaches you little that is new. A small model that predicts whether a molecule dissolves in water puts you in the exact area where frontier AI still struggles.
Skills worth building:
- Python, plus enough statistics to judge whether a prediction is any good
- Basic organic chemistry: functional groups and how structure links to behaviour
- Reading molecules written as text, and handling them with a chemistry library
- Reading a research paper's methods section without panic
If you are still deciding on the ML route itself, start with how to become an AI/ML engineer after 12th. For the broader question of which skills stay valuable, read how an early-career engineer can stay ahead of AI.
How to start learning this for free today
You can begin with the published papers behind the benchmark and a small property-prediction project, which costs nothing.
Read OpenAI's page
Go through the MedChemBench announcement and note how each ability is scored.
Pick one linked paper
Choose one paper or patent the benchmark draws on. Read the abstract and look at the structure tables.
Learn to read a molecule as text
Chemists write molecules as short text strings called SMILES. For example, "CCO" is ethanol. Practise turning a few drawings into strings.
Predict one simple property
Install the free RDKit Python library (
pip install rdkit). Use it to calculate logP, a number for how oily or watery a molecule is, closely linked to whether it dissolves. Here is ethanol, the "CCO" string from the last step:from rdkit import Chem from rdkit.Chem import Crippen mol = Chem.MolFromSmiles("CCO") print(Crippen.MolLogP(mol)) # -0.0014A value near zero means ethanol is balanced between oily and watery. If the call returns nothing useful, check that the SMILES string loaded.
Compare with a general model
Ask a chatbot to estimate the same property. Note where it agrees with your calculation and where it guesses.
Open OpenAI's MedChemBench page today, pick one of the published papers it draws on, and write down in your own words which of the four abilities that problem tests.