In 64 hours, Claude designed 318,484 proteins for a team of investment analysts, and picked 180 of them to test.
Five people at Citrini Research, three of them with no science background, set out this summer to see whether they could make a drug candidate with AI. They budgeted $200,000. I wrote about it in my channel on Tuesday, and it’s one of the best field reports on AI in biology I’ve read this year.
The same day, OpenAI announced it has acquired Orchestra Bio. The two stories make more sense read together.
What the analysts actually did
They chose scorpion antivenom. Scorpion antivenoms still come from horses injected with small doses of venom. The team went for minibinders, small proteins designed to stick to one specific toxin.
Working at Anthropic’s offices on an internal research model, they wrote a 35-page prompt together with Claude: five toxins, design constraints, tools, selection rules. Claude ran more than 300 subagents. They read papers, worked with specialist protein models and ranked candidates. When a manufacturing question came up, Claude emailed a scientist at Twist Bioscience, waited for the answer and adjusted the designs.
Compute cost $49,997. Lab work at Twist and Adaptyv Bio cost close to $150,000.
Two labs tested 211 designs. 17 met the binding criteria, covering four of the five toxins. The moonshot, one protein for all five toxins, didn’t work, and Claude had warned them it probably wouldn’t.
The number I keep coming back to
Of the 60 designs aimed at a single toxin, 11 bound. Anthropic says that 18% is normal for this kind of campaign.
So the design step got very fast, and the share that survives contact with a lab did not move. Nothing in the computer told anyone which 11 it would be.
Binding a toxin is also a long way from neutralizing it. Anthropic is explicit that it has not tested whether these proteins stop the venom, so nobody has an antivenom. Cell tests come next, then animals, then years of trials.
Citrini’s investment conclusion is to be long experiments. Cheap design sends more molecules into the same expensive labs, so total costs may not fall at all.
Where OpenAI is placing its bet
Orchestra Bio builds a platform where people and AI agents run drug programs together. Agents track studies and timelines, and scientists and program leads make the decisions. OpenAI is combining it with Rosalind, its models and tools for life sciences.
OpenAI’s own example is very ordinary. A failed manufacturing batch delays the start of a clinical trial. Agents find what else that affects, gather the information and lay out options.
The example is about coordination between departments and partners, on systems that were never connected. OpenAI puts US drug development at 10 to 15 years from target to approval, and a lot of that time sits in exactly this work.
A founder Citrini spoke with made a similar point. When he took promising molecules to pharma executives early in his career, they already had more programs than they could afford to develop.
What I see from the pharma side
In pharma, one product’s story tends to live in a CRM, a data warehouse, field reporting and a set of partner portals, each bought by a different team at a different time.
That’s the layer OpenAI just bought into. Molecules are getting cheap to design. Keeping a program moving across those teams and vendors is still slow and expensive, and very little of it is automated.
It reminds me of software. Code gets written faster than ever, and the bottleneck moved to architecture, testing and checking whether it works in a real business. The difference is how fast you find out. A bug can show up in seconds. Whether a drug works in people takes years to prove.
What I’d watch
Two things.
First, who keeps the experiment data. Insilico, which I wrote about last week, built its business on its own molecules and a large base of experiments. A design, test and learn loop only compounds if the results feed the next round, failures included.
Second, who owns the program layer inside big pharma. If agents start tracking studies, batches and timelines, whoever runs the system they work in sees the whole program.
The molecule is becoming the fast part, and the money will follow whoever speeds up the slow part.