OpenAI President Greg Brockman reposted a message that quickly drew attention across the technology and pharmaceutical sectors. On X, scientist Andrew Aiginin said GPT-6 Astra had beaten all frontier models in an independent benchmark and had become the strongest AI system for antibody developability prediction.
According to Aiginin, Astra did more than place first. He also said the model built an interactive visualization page for a complex antibody mechanism in roughly one hour.
GPT-6 Astra scored 37.98 on the DDD Benchmark antibody module
The reported result places GPT-6 Astra at 37.98 on the antibody developability module of the DDD Benchmark. The article says that module combines six different kinds of antibody experimental data, and that Astra finished ahead of every other model included in the test.
AI scholar Rim Shayakhmetov reacted to the result on X, writing: "What are the typical scores of specialized models on this task? This is so interesting — you rarely see frontier large language models shine this brightly on drug discovery benchmarks without relying on external specialized tools."
The article presents that performance as evidence that general-purpose large models are no longer limited to helping with coding or literature review, but are entering harder and less quantifiable stages of biomedical R&D.
Binding is only the start; developability is the harder hurdle
Aiginin summarized the challenge in a separate post: "Finding an antibody that binds is only the beginning. Will it aggregate? Will it stay stable? Will it actually behave like a drug?"
That framing points to a central problem in antibody development. The article says that much of the field has long focused on binding strength — whether an antibody can recognize and attach to a target such as a virus or cancer cell. Earlier AI systems, including AlphaFold in the article’s comparison, have already shown strong performance on prediction problems tied to binding. But a binder is not automatically a drug.
The piece lists the practical questions that matter in development:
- whether the antibody will aggregate;
- whether its structure will remain stable;
- whether it will behave like a real therapeutic.
Those properties are grouped under antibody developability. The article describes this stage as a common failure point, where many promising antibody candidates perform well in the lab but do not survive downstream druggability assessment, consuming large amounts of money and time in the process.
Insilico Medicine paper lays out the benchmark philosophy
The benchmark result did not appear in isolation. The article links it to a 21-page arXiv paper from Insilico Medicine titled "Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis."
The source provided in the article includes the PDF at https://arxiv.org/pdf/2608.18940, while the reference section also lists https://arxiv.org/html/2608.18940.
According to the report, the paper gives the first detailed look at how strict the DDD Benchmark is and shows how the team built an evaluation framework in small-molecule chemistry. Bogdan A. Zagribelnyy, PhD, said Astra topped the benchmark platform built by his team. While the social media discussion centered on large-molecule antibodies, the paper itself focuses on the depth of the evaluation setup in small-molecule chemistry.
The paper focuses on single-step retrosynthesis and rethinks evaluation
The study centers on single-step retrosynthesis, described in the article as the process of starting with a target drug molecule and working backward to identify accessible chemical building blocks that could produce it.
Older evaluation methods often worked by giving a model a target molecule and then checking whether the generated synthetic route exactly matched a route in a patent database such as USPTO. The article argues that this is poorly aligned with real chemistry because synthesis is naturally a one-to-many problem, while a single-answer scoring system punishes valid but novel routes.
To address that, the Insilico team introduced a layered evaluation system.
URSA-expert-2026 benchmark
The paper’s URSA-expert-2026 benchmark contains 100 newly generated molecules whose synthetic accessibility was manually confirmed by top human chemistry experts. The article says the set is fully isolated from public training data, reducing the chance that a model can succeed by memorization.
ChemCensor scoring system
Instead of rewarding a model for reproducing patented routes, the team developed the ChemCensor framework. It scores generated reactions based on low-level chemical rules, including reaction-center mapping and functional-group compatibility, to judge whether the output is chemically plausible.
Under that stricter test, the article says conventional large language models struggled in early Top-1 evaluations. Even GPT-5.5 and Gemini 3.1 Pro did not surpass specialized chemistry systems such as MHNreact and LocalRetro.
Three methods used to unlock stronger chemistry performance
The paper’s answer was not to discard general-purpose models, but to change how they were prompted and trained. The article highlights three methods.
Top-K prompting
Because retrosynthesis is a one-to-many task, the team explicitly asked models to generate 15 different answers rather than one. The article says that simply switching to a Top-K setup led to broad gains in chemical plausibility and diversity across major models, including GPT-5.5 and the Claude Opus series.
Gemini 3.1 Pro, in particular, saw a large jump in Top-15 mode and became the strongest LLM baseline in the paper’s description. The article interprets this as evidence that the knowledge was already there, but the earlier prompting format limited performance.
A dataset of roughly 45.6 million validated reactions
To train a chemistry-constraint-aligned language model called C3LM, the team used a virtual synthesis engine together with ChemCensor to build a large dataset of around 45.6 million validated real chemical reactions. The article refers to this corpus as CREED-CCV-2+USPTO-XL.
Novelty reward in reinforcement learning
During C3LM fine-tuning, the researchers used the GRPO reinforcement learning algorithm. The reward design was narrow and demanding: a route had to satisfy ChemCensor’s plausibility criteria, and routes that were chemically sound while also absent from the 45 million-reaction training set earned additional reward.
After that training process, C3LM beat traditional specialized SSRS models including MHNreact in blind testing on URSA-expert-2026. The article adds that the data analysis showed C3LM and other foundation models explored reaction spaces complementary to those covered by traditional models, producing many routes those older systems did not reach.
From C3LM to GPT-6 Astra, the discussion shifts to large-molecule prediction
The article ties that paper back to GPT-6 Astra’s score in antibody developability. Its argument is straightforward: in small-molecule synthesis, the Insilico team needed 45 million data points, reinforcement learning alignment, and Top-K prompting to push large models past specialized systems. Yet in a higher-dimensional large-molecule developability task, GPT-6 Astra reached the top of the same DDD Benchmark framework without external tools and posted a score of 37.98.
The piece contrasts that with older workflows in which antibody stability work often required dedicated teams and specialized 3D graph neural networks. It presents Astra as a general model showing competence across physics, chemistry, and biology tasks, and treats the benchmark result as a notable post-AlphaFold shift.
Aiginin also wrote: "By the way, the interactive visualization page showing how the antibody actually works was also built with Astra in about 1 hour." The article compares that with a more conventional pharmaceutical workflow that, in its description, could take two to three weeks to deliver a similar system.
References and source details
The article lists the following references:
- https://arxiv.org/html/2608.18940
- https://x.com/gdb/status/2098167375342239957
- https://x.com/andrewaiginin/status/2098078245350518884
The MarsBit page says the piece originated from the WeChat public account "新智元," with ASI启示录 credited as the author and Aeneas 大卫 as editor. MarsBit shows the publication time as 2026-09-11T04:23:10.000Z.