OpenAI buys Orchestra Bio: why AI lab deals in biotech will stay narrow

OpenAI bought Orchestra Bio on Tuesday, six months after Anthropic paid about $400M for Coefficient Bio. We're already seeing a lot of VCs betting on the large AI labs as a new exit channel for their startups. I'm not so sure this is sustainable though.

Let's first look at Orchestra: it spelled out its job in a June blog post it has since taken offline. The concept was simple and intriguing: a biotech can now take a program to IND without any lab space of its own, so the work runs at CROs and CDMOs and someone has to steer it.

Orchestra's agent pulls deliverables, due dates, payment milestones and risks out of the statement of work. The team forwards the CRO's emails and meeting notes, and the agent tracks what is slipping, whether tumors in an in vivo study grow late or a manufacturing run gets delayed.

OpenAI is folding all of this into Rosalind, next to the GPT-Rosalind model it launched in April and Rosalind Workbench, which helps researchers plan experiments. Orchestra came out of stealth last year with $20M, and terms of the deal weren't disclosed.

It is important to know that both big AI lab deals were narrow so far. Anthropic's $400M went to fewer than ten computational biologists from Genentech, with no product on the market yet. Orchestra fills one gap in Rosalind, the operational agentic workflows around a program. Where Anthropic needed that kind of operations work, it so far went with partners: Veeva's AI agents use Claude, and Anthropic pitches a Medidata connector for catching enrollment and site problems early. For its protein binder study in August, it had Adaptyv Bio and Twist Bioscience run the lab tests. That said, it's too early to tell what Anthropic's own new wet lab will change.

For now I expect lab deals to stay this targeted, either a team the lab can't hire one by one or a workflow it can't get through a partner. Claude Science and Rosalind Workbench already cover ground that many AI research tool startups sell into, so for many of them the lab may show up as a competitor first instead of an acquirer.