GPT-6.2 Astra Scores 84.5% on OpenAI’s New MedChemBench

OpenAI introduced MedChemBench on Friday launching a 99-problem medicinal chemistry benchmark where GPT-6.2 Astra achieved 84.5% of the available score using the xhigh reasoning effort setting, unite.ai reported.

The benchmark tests whether AI models can link molecular structures to experimental findings. At the xhigh setting, GPT-6 Astra scored 83.3%, GPT-6.1 Sol reached 82.0%, and GPT-5.6 Sol recorded 58.2%, unite.ai reported.

Medicinal chemistry presents a core balancing act. A compound often displays high potency against a disease-related protein while suffering from poor absorption or rapid degradation.

Model Scores at Xhigh Setting
Model Score (%)
GPT-6.2 Astra 84.5
GPT-6 Astra 83.3
GPT-6.1 Sol 82.0
GPT-5.6 Sol 58.2

The technical paper details a retrospective task set drawn from 117 articles and patents. It evaluates 761 distinct molecular structures across preclinical drug development tasks, unite.ai reported.

These tasks stop at a research decision point and provide biological background and early experimental findings. Later results are withheld to establish answer keys.

Seven categories divide the tasks, featuring 48 problems for candidate selection and 26 for structure-activity relationship prediction. Smaller groups cover ADME, molecular design, safety, retrosynthesis, and mechanism of action.

GPT-6.2 Astra reached 84.5% overall, but subcomponent scores show stark contrasts. GPT-6 Astra achieved 95.37% for molecular structure and 91.05% for scientific reasoning at xhigh, unite.ai reported.

Conversely, candidate selection and numerical predictions lagged significantly at 50.42% and 47.02%. In a reconstruction test involving an 18-atom ring, GPT-6 Astra produced correct structures in all ten trials, while GPT-5.6 Sol and GPT-5.6 Luna failed entirely, unite.ai reported.

OpenAI said that the benchmark does not prove whether models can discover new drugs in physical laboratories. Each response relied on a single attempt without tools, feedback, or prospective experimentation, unite.ai reported.

Proposed expansions include prospective tasks, expert evaluation of alternative answers, and agent-based frameworks that allow multi-step verification, unite.ai reported.

The benchmark evaluates 99 preclinical medicinal chemistry problems covering molecular structures, experimental findings, and candidate selection.

GPT-6.2 Astra achieved the highest reported score of 84.5% using the xhigh reasoning effort setting on November 20, 2026, unite.ai reported.

OpenAI, founded in 2015 as a nonprofit with Elon Musk and Sam Altman as co-chairs, released ChatGPT in November 2022, which by September 2026 had become the fifth-most-visited website globally. A 2025 restructuring converted its for-profit subsidiary into OpenAI Group PBC, a public benefit corporation 26% owned by the nonprofit OpenAI Foundation, after a March 2026 funding round valued the company at US$852 billion.

The technical paper assessed 761 distinct molecular structures gathered from 117 articles and patents, unite.ai reported.