AI Drug Discovery Money Concentrates as Anthropic Enters Biotech

A month that changed who is in the race

In October 2026, the AI drug discovery business got more crowded and also more top-heavy. The two leading frontier-model labs both moved into biotech. Big pharma kept signing platform deals worth billions in potential payments. Investors put large sums into a fairly small group of protein-design companies. The pattern across all of this is clear. Money and attention are going to a few groups that can link models, wet labs and clinical pipelines. Better structure prediction alone no longer earns a premium. At the same time, the sector still has to prove that a drug discovered mainly by AI can win approval.

The frontier labs arrive

Anthropic made the biggest headline. On October 27 it said it had acquired Mithril Medicine, a biotech run by one person, founder Cody Tranbarger, who used Claude to coordinate contracts and experiments across 11 outside laboratories while working on rare muscular dystrophies.8 According to the coverage, Mithril took a drug candidate from concept to successful mouse studies in under 18 months, and human trials could start as early as late 2028.8 Anthropic also described results from six partners: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group.9

The partner claims are specific. BigHat reported antibody fragments that bind more than ten times more tightly. Divergence Labs produced about 35,000 antibody designs for food allergies in a month. Generate said Claude helped it write CAR-T genetic instructions about 70% shorter than usual.9 Enveda said review of a 1,061-page manufacturing document dropped from more than 25 hours to under three, and that Claude found seven problems human reviewers had missed.8 Both reports stress that the work is preclinical and that the data come from the companies themselves.89 One report links the push to Anthropic's IPO filing. On this view, named pharma partners give public-market investors a revenue story beyond chatbots and coding.9

OpenAI moved first with its purchase of Orchestra Bio, whose AI agents run clinical-trial operations. The price was not disclosed.6 Reports disagree on the date: one puts the deal on October 20 and another, citing Cryptonomist, on October 22.86 The two labs are aiming at different stages, though. OpenAI's deal targets the trial operations that dominate development costs. Anthropic is placing its model in the design-and-paperwork work of early discovery. Both are building a presence in life sciences, but they disagree on where general-purpose AI adds the most value first.

Pharma keeps writing big but conditional checks

The month's largest platform deal was AstraZeneca's alliance with insitro. It focuses on MASH, a fatty liver disease, and on cardiomyopathies, using insitro's "Virtual Human" platform, which combines machine learning with human genetics and functional genomics.3 Reports agree on a $40 million upfront payment. They differ on the headline value: pharmaphorum called it a $1.5 billion partnership, with $1.46 billion in milestones, while Bloomberg put it at up to $1.86 billion, apparently counting additional components beyond the milestones.35 Either way, almost all of the money depends on future milestones. That has become the standard shape of these deals.

Generate Biomedicines' collaboration with Zealand Pharma on metabolic-disease proteins follows the same pattern. Generate could earn close to $1.2 billion in milestones on two initial programs, but none of that is guaranteed.17 Elsewhere, Astellas agreed a $400 million computational discovery deal with AI Proteins, and Eli Lilly committed up to $1.1 billion to a tissue-targeted lipid nanoparticle partnership with Metis TechBio.4 Boehringer Ingelheim and Microsoft announced work on "digital patient twins" for lung disease, but gave no patient numbers, accuracy figures or timeline.6

The infrastructure side is also growing. Danaher said it will open an autonomous AI lab at Abcam, expected to reach full capacity in early 2027. It is designed to run a continuous design-make-test-learn loop for antibodies and other reagents, with up to eight times faster progress and ten times more reagents produced each year.1 Biohub's Virtual Biology Initiative grew from $500 million to $1.8 billion with partners including Google DeepMind, Isomorphic, Meta, the NIH and the Department of Energy, although much of that value is data and computing rather than cash.6

Protein design funding: more money, fewer winners

The strongest case that AI protein design has become a late-stage market comes from a tally covering January 1 to September 24, 2026. In that period the sector raised about $1.03 billion across 23 transactions, roughly four times the comparable 2025 figure, with about 28% more deals.11 Most of the growth came from a few very large rounds. Chai Discovery's $400 million Series C alone accounted for about 39% of the money, at a reported $3.8 billion valuation, close to triple its valuation seven months earlier.11 Without rounds above $50 million, funding rose only about 26% and stayed well below 2024 levels.11

Antibody and binder design is now the dominant category, with nine deals and about $724 million, roughly 70% of the quantified money.11 Besides Chai, the backed companies include Aureka, which raised twice; Infinimmune and BigHat, at $75 million each; and Galux in South Korea.11 Meanwhile only two first-time financings were recorded, down from five in each of the two previous years.11 In plain terms, investors are backing established platforms and avoiding new founders.

The flow continued into autumn. September brought Enveda's $311 million Series E, Anew Labs' $290 million first external round and Basecamp Research's $140 million Series C.16 Beijing-based BioGeometry said on October 28 that it had raised more than $100 million across Series A rounds for its protein-design models.12 China is a growing part of this picture. Matwings Technology, a protein foundation-model company, closed a Series A+ of more than RMB 200 million in March and another round in July.11 Xaira Therapeutics, which launched in 2024 with more than $1 billion, revealed its X-Design antibody platform and first pipeline details this month.2

Structure prediction becomes table stakes

Isomorphic Labs, Alphabet's AlphaFold spin-off, shows both the promise and the strain in the field. It raised $2.1 billion in May in a round led by Thrive Capital.13 By October it was reportedly seeking new money at a valuation of at least $40 billion.28 Its pitch now centers on IsoDDE, a proprietary engine the company says more than doubles AlphaFold 3's accuracy on a hard protein-ligand generalization benchmark. The company also says IsoDDE predicts binding strength better than physics-based methods and can find new binding pockets from sequence alone.2125 Unlike AlphaFold 3, IsoDDE is fully proprietary.26

The Boston Globe reported that outside researchers find the technical paper promising.22 The same reporting noted that independent studies have found AlphaFold-style models can memorize training data and miss binding sites, even though AlphaFold did better than its rivals in one University of Basel stress test.22 Isomorphic has not dosed a patient. Its first trials, originally targeted for the end of 2025, are now expected by the end of 2026.262 With weeks left in the year, that deadline is a real test.

In our reading, the market has already priced in what one analysis calls the commoditization of the prediction layer. With AlphaFold 3 code, Boltz and Chai models widely available, value is shifting to the full discovery loop.24 Google now offers AlphaFold 3 for deployment on its cloud platform, which supports the point.23 Isomorphic's move "beyond AlphaFold" and Danaher's closed-loop lab both respond to the same pressure.221

The clinic is still the only scoreboard

Coverage agrees that clinical results now matter more than the technology itself. Iambic priced a $150 million Nasdaq IPO at $16 a share on October 14, after raising $462 million privately, and investors are judging it on its HER2 drug IAM1363.104 Insilico's rentosertib, chosen by AI at both the target and molecule stage, began dosing in a 320-patient Phase 3 trial for lung scarring (IPF) in China in September.6 Generate's anti-TSLP antibody is in replicate Phase 3 asthma trials.10

The setbacks are real too. Neumora's lead drug failed two Phase 3 studies, which shows that machine learning does not remove clinical risk.10 No drug discovered mainly by AI had been approved as of October. The closest candidate, Takeda's zasocitinib, designed with Schrödinger's physics-based platform, faces an FDA decision expected in early 2027.2

The overall picture is a sector that is getting richer faster than it is producing proof. Valuations like Isomorphic's and Chai's assume the design loop will cut failure rates, not just speed up discovery. Over the next 18 months, Phase 3 readouts will matter far more than the next benchmark. Until then, the most reliable gains are the ones Anthropic's partners reported: faster paperwork, tighter antibody binding and shorter regulatory drafts. They save money, but they do not yet show that AI picks drugs that work.