GPT-6 ranks first, which is the most stunning breakthrough after AlphaFold, and has become the most powerful AI for antibody prediction.
The biggest shock since AlphaFold has arrived.
Just now, a message reposted by OpenAI President Greg Brockman is enough to ignite the entire technology and pharmaceutical sectors.
Renowned scientist Andrew Aiginin posted an exciting announcement on X —
In rigorous independent benchmark tests, GPT-6 Astra has just outperformed all cutting-edge models and become the most powerful AI for antibody developability prediction!
Not only did it top the leaderboard, Astra also directly generated a complex interactive visualization page for antibody mechanisms in just 1 hour.
For a long time, there has been a consensus in the industry: "AI for Science must use dedicated models to solve dedicated problems."
For example, AlphaFold corresponds to protein folding, while LLMs are usually only regarded as auxiliary tools for coding and literature review.
But today, GPT-6 Astra has torn a huge hole in this consensus!
Relying on general intelligence, it has directly targeted the most difficult-to-quantify and most troublesome preclinical bottleneck in biomedical R&D.
None of this happened by chance. The top AI pharmaceutical team Insilico Medicine, which built the test benchmark, released a 21-page landmark paper on arXiv.
Title: Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis
Paper: https://arxiv.org/pdf/2608.18940
This paper not only reveals for the first time the rigor of the DDD Benchmark used to test GPT-6 Astra, but also shows us that from small-molecule chemical synthesis to large-molecule antibody prediction, general large models are taking over the hardest core boundaries of science in an all-round way!
The scientific paradigm shift after AlphaFold may have only just begun.
37.98 Points! GPT-6 Tops the Authoritative "DDD Benchmark"
Actual test results show that after accessing the antibody developability test module in the DDD Benchmark (which integrates 6 different sets of antibody experimental data), GPT-6 Astra achieved an astonishing score of 37.98, crushing all tested models.
AI scholar Rim Shayakhmetov made an incredulous exclamation:
"What is the usual score of dedicated models on this task? This result is so interesting. It is very rare to see a cutting-edge large language model perform so brilliantly in drug discovery benchmarks 'without relying on external dedicated tools'!"
Without professional bioinformatics plugins and without dedicated fine-tuning on antibody data, a general large model has directly hit the core of antibody druggability, which is extremely exciting.
Discovering Antibodies Is Only the Beginning, "Druggability" Is the Real Valley of Death
To understand the value of GPT-6 Astra's achievement, we must first understand a cruel reality in the pharmaceutical industry.
Andrew wrote a soul-stirring sentence: "Finding an antibody that can bind is only the beginning. Will it aggregate? Can it remain stable? Can it really function like a drug?"
This sentence speaks to the blood, sweat and tears of countless new drug R&D personnel around the world.
In the past, when we talked about antibody drug R&D, everyone's top concern was "binding affinity".
In layman's terms, it is whether this antibody can accurately recognize and bind to the target on a virus or cancer cell, just like a key inserted into a lock.
Early AI models (such as AlphaFold) have already done an excellent job at "predicting binding".
However, being able to bind ≠ being able to become a drug.
This is just like falling in love. Two people take a fancy to each other and fall in love at first sight, which is only the first step. Can you walk down the aisle together? Can you withstand the test of daily necessities?
In drug R&D, this "test of daily necessities" is called "druggability".
In reality, countless "perfect antibodies" that seem to have extremely strong binding affinity will expose fatal flaws once they enter the real physiological environment or industrial production process:
- Will it aggregate into clumps?
- Is its structure stable?
- Can it really perform like a genuine "drug"?
These properties are collectively referred to as antibody developability.
This is the "Valley of Death" for the entire biopharmaceutical industry.
It has led to an extremely tragic situation: every year, thousands of antibody molecules around the world perform excellently in the laboratory, but eventually fall at the "druggability" evaluation stage.
Pharmaceutical companies have burned billions of dollars and spent years or even more than a decade on this, only to end up with nothing.
Today, GPT-6 Astra announces that it can predict these druggability characteristics more accurately than any other cutting-edge model in the world.
Unveiling the Touchstone: The "Hell-Level" Benchmark Test That GPT-6 Astra Faced
Some people may ask: Could this just be another "leaderboard-chasing" achievement?
Absolutely not.
Dr. Bogdan A. Zagribelnyy pointed out that the platform where Astra topped the list is the DDD Benchmark they built.
In the paper they published, we see how hardcore this test system is.
Although the tweet discusses "large-molecule antibodies", this paper demonstrates the terrifying depth of the team's establishment of evaluation standards in the field of "small-molecule chemistry". Understanding this background will make you realize how incredible Astra's championship is.
In the paper, the research team focused on another century-old problem in the pharmaceutical industry: single-step retrosynthesis.
Simply put, retrosynthesis is the "reverse deduction method in the chemistry field" — given a target drug molecule, you need to reverse deduce which existing, cheap chemical raw materials can be used to synthesize it.
How were AI evaluations done in the past?
Give the AI a target molecule, and the AI outputs a synthesis path. If it is exactly the same as the one in the existing patent database (such as USPTO), it gets a full score of 100.
But this is almost unworkable in the real world. Because synthesis paths are inherently "one-to-many", the traditional "single standard answer" evaluation method greatly limits the ability of AI to explore the unknown.
To this end, the Insilico team built a ridiculously rigorous test environment:
1. URSA-expert-2026 Benchmark: This is a new molecular dataset containing 100 machine-generated molecules whose synthetic accessibility has been manually confirmed by top human chemistry experts. It is completely isolated from any public training data, completely eliminating the possibility of the model "memorizing answers".
2. ChemCensor Scoring System: Instead of checking whether the model perfectly reproduces patents, they developed the ChemCensor framework to judge whether the reactions generated by the model have real chemical rationality, starting from the most underlying chemical and physical rules such as "reaction center mapping" and "functional group compatibility".
It was under this kind of "peel back the layers to see the core" assessment of physical and chemical laws that almost all traditional LLMs failed.
In early Top-1 tests, even as powerful as GPT-5.5 and Gemini 3.1 Pro, their performance still could not surpass the top dedicated traditional chemical models (such as MHNreact, LocalRetro).
The "greater computing power brings better results" general LLM seems to have hit a wall.
45 Million Data Points, How LLM Was "Forced to Develop" Scientific Intuition?
General LLMs cannot beat dedicated models, so what should we do?
The answer given by the paper shocked the industry — do not change the model, change the way to stimulate the potential of the general model.
The research team proposed three killer moves in the paper to completely unlock the seal of large models in the chemistry field:
First Move: The Top-K Prompt Paradigm
Since retrosynthesis is "one-to-many", don't let the LLM guess only once.
The team explicitly stated in the Prompt: "Give me 15 different answers".
A miracle happened. Just by switching to the Top-K mode, almost all large models (GPT-5.5, Claude Opus series) achieved explosive growth in chemical rationality and diversity.
The performance of Gemini 3.1 Pro skyrocketed in the Top-15 mode, directly becoming the strongest LLM baseline.
This proves that large models have scientific knowledge, but the previous way of asking questions was wrong.
Second Move: Build a Super Large Verification Dataset of 45.6 Million Scale (CREED-CCV-2+USPTO-XL)
In order to train a language model specifically aligned with chemical constraints (C3LM), the team spared no computational power, and used a virtual synthesis engine and the ChemCensor framework to build a super-large dataset containing about 45.6 million verified real chemical reactions.
Third Move: "Novelty" Reward in Reinforcement Learning
When fine-tuning the C3LM model, they introduced the GRPO reinforcement learning algorithm.
The reward function is extremely ingenious: the synthesis path given by the model must not only meet the chemical rationality of ChemCensor, but if the model can come up with a completely new reaction that does not even exist in the 45 million training set but is still reasonable, it will get an additional huge reward.
What was the result?
The C3LM model trained with this "hell-level" training finally defeated all traditional dedicated SSRS models such as MHNreact in the URSA-expert-2026 blind test!
Data analysis shows that C3LM and other basic large models can explore a reaction space completely complementary to traditional models, generating a large number of unique paths that traditional models cannot think of.
The core conclusion of the paper is: Large language models can not only process natural language, but as long as they are given the right environment and incentives, they can fully master the most profound chemical bonds, functional groups and synthesis logic!
Dimensionality Reduction Strike: From C3LM to GPT-6 Astra, the Full Takeover of General Intelligence
If you understand the above paper, you will understand why Dr. Bogdan linked GPT-6 Astra's achievement to this paper.
In this paper, the Insilico team went to great lengths to fine-tune with 45 million data points, align with reinforcement learning, and guide with Top-K prompts, finally making the large model beat dedicated models in small-molecule chemical synthesis.
However, when the test track switches to the more complex, higher-dimensional large-molecule developability prediction, GPT-6 Astra even topped the list directly under the same DDD Benchmark test system without using external tools, scoring 37.98 points!
This is absolutely a dimensionality reduction strike.
In the past, to solve the problem of antibody stability, a professional team needed to design a dedicated 3D graph neural network.
Now, GPT-6 Astra's general world model seems to have internalized the underlying logic of physics, chemistry and biology.
After reading the vast number of papers, patents and experimental data in human history, and even learning the physical laws of the microscopic world through multimodal learning, it can "intuitively" judge why a protein molecule aggregates in solution just like a human expert.
In addition, the engineering efficiency demonstrated by Astra is equally daunting.
Andrew Aiginin mentioned: "By the way, this interactive visualization page demonstrating how antibodies actually work was also built with Astra in about 1 hour."
In traditional pharmaceutical companies, building such a system would take at least two or three weeks for the entire process.
Now, it can be solved in 1 hour by one person using a general large model.
The Real Paradigm Shift After AlphaFold
GPT-6's top score on the antibody benchmark declares that general large models have shown the embryonic form of AGI, achieving substantial dominance in the most top-level intellectual activities of human beings.
This top spot on the leaderboard announces the arrival of this new era —