Anthropic's "not a new model" launch cost Schrodinger 8.3%. The fight moved to plumbing

The equity reaction is the part worth reading, because there was nothing in this launch that a model-scale scoreboard would register. Anthropic said in plain terms that Claude Science is not a new AI model and not a more capable model for biology ^[13]. What it is: existing Claude models routed through connections to more than 60 scientific databases and specialized tools, able to carry a workflow across several steps ^[1], with a second agent auditing citations, figures and numbers before a human reads the output ^[2]. Eric Kauderer-Abrams, who runs life sciences at Anthropic ^[11], describes that as the old scientific division between the group doing the work and an independent group reviewing it ^[3], and notes the agents each carry their own context and stay fallible ^[5]. Coordination sits with Claude; structure prediction and sequence analysis stay with the tools built for them, reached in part through Nvidia's BioNeMo ecosystem ^[6].

The revenue disclosures deserve less deference than they usually get. A run rate of roughly $47 billion in late May 2026 against about $9 billion at the end of 2025 ^[8] is a 5.2x move in roughly five months ^[15]. The other headline figure, more than 1,000 customers at $1 million a year or more ^[9], sets a floor of about $1 billion, or some 2% of that run rate ^[16]. The cohort everyone quotes accounts for a small slice of the number everyone quotes, and where the rest comes from is not broken out.

The productivity claim has the same shape. Anthropic reports timelines cut roughly tenfold in some cases, including Stephen Francis' lab at UCSF, where researchers manually validated a Claude-generated glioma review and still finished in about a tenth of the previous time ^[17], a cut of about 90% in elapsed time ^[18]. Verification did not disappear; assembly did. The system cannot do the bench work that generates the data, and scientists must check results before consequential steps such as a regulatory submission ^[14]. A lab's accumulated institutional knowledge still sits with the people holding it ^[14].

For anyone scoring vendors, the questions reorder themselves. A benchmark position says nothing about whether a system can reach the data a given workflow needs, or who signs the output. The suppliers most exposed are the ones whose product is a single step in a chain that something else now coordinates ^[1]. The Forbes account frames the addressable gap as the distance between a demo that impresses a data-science team and a deployment a hospital or a bank will actually trust ^[12].