Anthropic reports Claude-designed minibinders against scorpion toxin

In August 2026, a team of analysts at Citrini Research submitted roughly 35 pages of prompts to an internal Anthropic research model, kicking off an exploratory protein-design experiment. According to a report from Cailian Press on October 20, the model ran more than 300 sub-agents overnight and drew on over 300 parallel GPUs to design, score and filter proteins. The experiment had a budget of $200,000, and three of its five participants had no scientific background.

Anthropic has since published the results. Claude generated candidate minibinders for scorpion neurotoxins, producing 318,484 designs spanning 553,052 design-toxin pairs. Of the 211 designs sent for laboratory testing, 17 met the predefined binding criteria, with binding tests carried out by two external contract research organizations. Citrini said the 211 designs were sent to Twist Bioscience for synthesis, and that the team spent about two more months waiting for and analyzing results because DNA synthesis, protein production and purification, sourcing usable toxins and setting up experimental controls all take time.

The two accounts differ on who led the work: the DEV Community report attributes the designs to the exploratory project Anthropic published, while Citrini's account emphasizes that its analyst team carried out the work using Anthropic's internal research model. The public results stop at the binding stage. Anthropic says it has not tested whether these molecules neutralize toxins in cells or living organisms, so they are not an antivenom. Citrini Research argues that as AI shortens the path from research idea to experimental plan, the bottleneck shifts to testing capacity and laboratory resources, with contract research organizations potentially playing a role akin to cloud infrastructure.