A CRISPR array is a stretch of DNA where short repeats sit at even spacings, like sleepers on a railway line; bacteria use it to remember viruses. On 23 September Anthropic said its model found a new one nobody had described, in a family of enzymes where nobody was looking.

Anthropic opened a life sciences group with a wet lab in the Bay Area and published its first finding. Roughly 950 Claude agents ran for 21 hours, spending 210 million tokens, scanning about 200,000 reverse transcriptases, enzymes that copy RNA back into DNA, and narrowing 3,500 candidates down to 20 systems. One pairs a reverse transcriptase with a neighbouring partner gene and a long array of evenly spaced repeats, which is the shape of a CRISPR system. Anthropic calls it array-associated reverse transcriptase, ART. What it does is unknown. Every laboratory experiment is run by human scientists; the model searched the genomic data, proposed candidates and helped read the results. Feng Zhang, who built the first CRISPR gene editing tools, called it genuinely intriguing.

Why this one is different

Models have been reading biology for years, and labs have used them to rank candidates somebody else chose. The unit here is different: a search nobody directed at a target, run long enough and wide enough that the interesting thing was found rather than confirmed. Twenty-one hours of agent time over two hundred thousand enzymes is a literature review no group of people would attempt, because the expected return per enzyme is close to zero. That is precisely the shape of work where a cheap tireless reader beats a careful one.

Nobody told it what to look for. That is the result.

How we got here

1. 8 Sep 2026DeepMind publishes AlphaGenome Atlas, predictions for every single letter change in human DNA.
2. 7 Sep 2026Six independent ageing clocks agree about an AI designed drug.
3. 17 Sep 2026Anthropic reports 26% of its own model research is now led by Claude.
4. 23 Sep 2026950 agents, 21 hours, and a CRISPR-like system with no known function.

What it does and does not mean

A structure is not a function. ART looks like CRISPR in its arrangement and may do something entirely unlike it, or nothing useful at all, and until the lab work is done and published this is a computational finding with a good pedigree and no result. The share prices that moved on the day moved on a resemblance. What the week does settle is a working method: the lab that measured how much of its own engineering the model now leads has started pointing the same machinery at somebody else's field, with humans holding the pipettes and the model holding the search.