GPT-6 Astra Tops Antibody Developability Benchmark, OpenAI President Says

On September 10, 2026, OpenAI president Greg Brockman reposted a message from Andrew Aiginin on X announcing that GPT-6 Astra had reached the top of the antibody developability prediction leaderboard on the DDD benchmark developed by Insilico Medicine, scoring 37.98 and beating all evaluated frontier models. Aiginin also said Astra built an interactive visualization page explaining antibody mechanisms in about an hour.

The benchmark's antibody developability module combines experimental data from six antibody datasets and assesses properties related to drug developability, such as aggregation tendency and stability. In this run, GPT-6 Astra used no external specialized tools and received no antibody-specific fine-tuning.

Insilico Medicine posted a related 21-page paper on arXiv the same day, titled Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis. The paper describes the URSA-expert-2026 dataset, which contains 100 new molecules validated by expert chemists for synthesizability, and the ChemCensor scoring system, which evaluates chemical plausibility through reaction center mapping and functional group compatibility.

The paper also details methods including Top-K prompting to generate multiple candidate solutions, training a chemistry-constrained language model called C3LM on about 45.6 million validated chemical reactions, and adding a novelty reward in reinforcement learning. The resulting C3LM model outperformed traditional specialized models on single-step retrosynthesis tasks.

In his post, Aiginin noted that finding a binding antibody is only the beginning, adding that researchers also need to determine whether it aggregates, remains stable, and behaves like an actual drug. GPT-6 Astra's result is being viewed as a demonstration of general-purpose models' capabilities in biology-related drug research.