The Clock
McKinsey published a chart this week that turns one of the grander promises of artificial intelligence into something considerably more useful: a calendar. Across a group of publicly disclosed AI-enabled drug programs, companies have reached preclinical candidates and first-in-human trials much faster than the timelines associated with conventional large-pharma development. The comparison deserves some caution from the start. These are selected programs that survived long enough and worked well enough for companies to talk about them, while the conventional benchmark also captures some of the organizational drag inside large pharmaceutical companies: portfolio reviews, sequential handoffs, competing priorities and the usual machinery of a very large organization.
Even after allowing for that, the compression is difficult to dismiss. McKinsey puts the traditional trip from an initial drug hit to a preclinical candidate at roughly 32 to 50 months, followed by another 16 to 20 months before the drug reaches its first human trial. Schrödinger nominated SGR-1505 as a development candidate about ten months after beginning its MALT1 program. Recursion and Exscientia reached a candidate with EXS21546 in nine. Generate:Biomedicines moved one program from hit to first-in-human testing in 17 months; Iambic and Absci have disclosed timelines around two years.
Those numbers shouldn't be treated as though every clock started from the same place. Schrödinger was working against MALT1, an established oncology target, and its achievement is primarily a story about computational chemistry. Recursion's REC-1245 program began further upstream with a biological relationship uncovered through its maps of cellular data. Insilico Medicine went further still with rentosertib, using its platform to identify TNIK as a potential target for idiopathic pulmonary fibrosis and then designing a drug against it. A ten-month hit-to-candidate program on known biology and an eighteen-month target-to-candidate program on comparatively novel biology are both interesting; they aren't the same accomplishment.
In The Age of Therapeutic Abundance, we looked at a broader change already appearing across medicine: more powerful therapies are beginning to collide with the capacity of the healthcare system to manufacture, finance and deliver them. AI drug discovery sits considerably further upstream. Faster candidate generation doesn't create therapeutic abundance by itself; it increases the number of experiments that can reach the places where biology, clinical development and eventually the healthcare system decide which ones survive.
Rentosertib has now made it farther through that sequence than most. Insilico moved from its AI-identified target to a preclinical candidate in about 18 months, and doctors in China dosed the first patients in a Phase III trial this month. The study will enroll roughly 320 people with idiopathic pulmonary fibrosis and follow them for 52 weeks, measuring whether the drug can slow the decline in lung function. The attractive part of the discovery curve has therefore met a rather less compressible calendar: researchers now have to give people the drug for a year and see what happens.
Eight Billion Possibilities
Schrödinger's MALT1 program offers the cleanest view of what has already changed. Researchers were looking for a small molecule that could inhibit MALT1, a protein involved in signaling pathways associated with several B-cell cancers, while satisfying the collection of other requirements that routinely kill otherwise promising molecules. A compound can bind tightly to the desired protein and still fail because it doesn't dissolve properly, can't cross the right membrane, disappears too quickly inside the body or binds somewhere else with unfortunate consequences.
Medicinal chemistry has always dealt with those tradeoffs through iteration. Researchers design compounds, make the most promising ones, test them, inspect the results and decide what to make next; the chemistry itself may be clever, but a considerable amount of calendar time disappears while ideas wait to become molecules and molecules wait to become data. Schrödinger pushed far more of that search into computation before asking the laboratory to make anything.
Its platform evaluated roughly 8.2 billion compounds while looking for promising chemical series. Researchers later assessed more than 5,000 proposed designs against the combination of potency, solubility, permeability and other properties they wanted, narrowing the field to 43 that satisfied their criteria closely enough to deserve further attention. Across the program, they synthesized 129 compounds, including 78 in the lead series, before selecting SGR-1505 as the development candidate about ten months after the work began.
Eight billion possibilities ultimately produced 129 physical compounds, which gets closer to the industrial significance of AI drug discovery than the familiar language about machines “inventing” medicines. The computer didn't eliminate medicinal chemistry, and it certainly didn't produce a perfect molecule on the first try. It allowed an enormous number of bad or mediocre ideas to die as calculations rather than compounds, preserving the laboratory for the much smaller number worth making.
The evidence for AI in drug discovery is strongest, at least for now, as an economics-of-search story. Schrödinger can consider billions of compounds without making billions of compounds; weak ideas can die earlier and promising ones can reach the laboratory faster. What hasn't been demonstrated is the more consequential claim that candidates produced by these systems are more likely to survive Phase II, Phase III and regulatory review. Getting to the answer faster has real value even when the answer is no. Getting to yes more often would be something else entirely.
The Search Moves Upstream
Recursion's REC-1245 program already stretches the search beyond chemistry. Its platform identified a previously unknown regulatory relationship involving RBM39 and CDK12, giving researchers a possible way to attack cancers with particular genomic vulnerabilities; the company then developed a molecule intended to degrade RBM39 and moved the program into IND-enabling work in less than 18 months after synthesizing roughly 200 compounds. The first patient entered a Phase I/II trial in December 2024.
On September 23, Anthropic reported an experiment that pushed the same idea much further into basic biology. Rather than asking Claude to optimize a molecule against a target researchers had already chosen, Anthropic's scientists gave it a high-level instruction to search a large genomic database for interesting examples of reverse transcriptases, enzymes that copy RNA into DNA. Roughly 950 Claude agents then spent 21 hours working through the data, using about 210 million tokens while gathering more than 200,000 reverse transcriptases and identifying roughly 3,500 candidate systems.
The agents narrowed those candidates to 20 for deeper analysis. While examining one unusual family, Claude noticed a repeating pattern of DNA beside the gene for a reverse transcriptase, compared its structure with known biological systems and searched the literature for evidence that anyone had described the combination before. Anthropic's scientists subsequently investigated the candidate in the laboratory and now describe it as an array-associated reverse transcriptase, or ART, a previously uncharacterized enzyme system found in bacteriophages with a repeat structure reminiscent of CRISPR.
The underlying reverse transcriptase itself had appeared in earlier research; Anthropic says Claude appears to be the first to recognize the neighboring DNA repeats and accessory protein as the defining features of a larger system. The work is also a preprint, Anthropic still doesn't know what ART actually does, and every wet-lab experiment was performed by human scientists. This is an interesting biological discovery, not a new gene-editing therapy discovered autonomously by a chatbot.
Still, it is a different kind of result from searching billions of candidate molecules. The useful object being searched was biology itself: hundreds of thousands of genomic sequences containing systems whose functions were poorly understood or completely unknown. Anthropic says an expert scientist might spend weeks or months carrying out the sort of analysis its agents compressed into roughly a day, and the agents weren't merely retrieving a previously documented answer. They were looking for anomalies, developing hypotheses about them and deciding which candidates merited scarce human attention.
The asymmetry that appeared in Schrödinger's chemistry program shows up again, just one level earlier. Claude could generate and discard thousands of biological leads before scientists chose which ones deserved experiments, much as computational chemistry can discard enormous numbers of molecules before researchers synthesize them. The scarce resource shifts from producing another hypothesis toward knowing which hypothesis deserves a piece of the physical world.
Biology Has a Different Clock
Anthropic's experiment complicates any clean line between fast computation and slow biology, because some of the work normally described as biological discovery can plainly be accelerated too. Reading literature, searching genomic neighborhoods, comparing protein families and noticing unusual patterns are intellectual search problems; agents can run more of those searches in parallel than a laboratory could ever staff with scientists. If that approach generalizes, the supply of plausible biological hypotheses could increase very quickly.
The answer to those hypotheses still comes from somewhere else. Anthropic's agents could identify ART in 21 hours, but they couldn't determine its actual function simply by thinking harder about the sequence. Scientists had to express proteins, run biochemical experiments and examine what the system did; months of additional work may follow before anyone knows whether ART is merely an interesting curiosity or the beginning of another useful biological tool.
Drug development encounters the same distinction with much higher stakes. A gene can appear repeatedly among people with a disease without causing the disease, while a protein that looks essential in a cell line can prove surprisingly optional inside a person. Even sophisticated experimental models simplify something. Mice aren't small humans, organoids aren't patients and cancer, fibrosis or neurodegeneration doesn't reorganize itself into tidy training data because researchers have acquired more GPUs.
McKinsey consequently finds AI considerably more mature in molecule design than in target identification, preclinical research or translation into humans. Anthropic's result suggests that target discovery and hypothesis generation may begin moving faster as well, although generating a promising biological explanation and demonstrating that it actually causes disease remain different jobs. The feedback arrives more slowly, contains more noise and often leaves several explanations alive at once; cells also take however long cells take to grow, while a chronic disease may need months before researchers can reliably see whether anything changed.
A beautifully optimized molecule aimed at the wrong mechanism remains a beautifully optimized failure, just as a clever genomic hypothesis can disappear after its first serious experiment.
Producing either one faster still has substantial economic value because researchers reach the bad news earlier and redirect time and capital toward another idea. The larger prize comes if better computational search also improves which hypotheses scientists choose to test in the first place.
The structure of pharmaceutical AI deals suggests that the industry understands the difference. McKinsey puts the headline value of major AI drug-discovery partnerships at $29.5 billion in 2025, but only about $600 million of that came as upfront payments. Most of the nominal value sits behind clinical and commercial milestones; among 38 deals signed during 2023 and 2024, less than one percent of the maximum milestone value had been paid by June 2026. Some of that simply reflects the years required for drug development, although the contracts also place most of the money beyond the point where a computationally attractive idea has to survive biology.
More Candidates, Same Patients
A discovery organization that can produce several credible candidates in the time once required for one eventually runs into a problem that another billion calculations cannot solve. Clinical trials need investigators, sites and patients; complicated therapies need specialized manufacturing; every additional program also needs the capital required to carry it through years of development. Those resources don't automatically multiply because the front end becomes more productive.
That is where the connection to The Age of Therapeutic Abundance becomes more direct. The report looked downstream, where better therapies increasingly collide with manufacturing capacity, specialist infrastructure, insurance rules and the ability to pay for treatment. Faster discovery can create a similar problem earlier in the process. Abundance at the front of a pipeline can produce congestion in the middle long before any of those molecules become abundant medicines.
Clinical trials make the mismatch easiest to see because one of their scarce resources is often a person. Two companies developing treatments for the same relatively narrow disease population may need many of the same investigators, hospitals and patients, while oncology programs can compete for people whose tumors carry very specific mutations. Faster candidate generation doesn't create another supply of eligible patients; it places more protocols in front of the same population.
Manufacturing presents another version as molecules become more complicated. A small molecule generally poses a different industrial problem from an antibody, cell therapy or antibody-drug conjugate; specialized facilities and expertise cannot always be conjured at the same pace as another computationally promising candidate. Development capital has its own limit as well. A company may be able to discover ten plausible drugs without possessing the money to carry ten drugs through clinical trials.
Anthropic's biology experiment contains a miniature version of the same problem. Nine hundred and fifty agents generated thousands of candidate systems in less than a day, but the laboratory couldn't run thousands of serious follow-up experiments at the same speed. Anthropic narrowed the field to 20 for deeper analysis and ultimately put physical resources behind a much smaller number. Once hypothesis generation becomes abundant, experimental attention becomes more valuable.
None of those constraints makes faster discovery less valuable. They change what the next valuable unit of capacity looks like. Once chemical and biological search become cheap enough, another billion virtual molecules or another thousand genomic hypotheses matter less than another reliable disease model, another high-quality clinical site, or a better way to decide which candidate deserves the scarce experiment.
The Test
Rentosertib is interesting because it has crossed from the part of the process where AI has already demonstrated speed into the part where the larger claim gets tested. Insilico didn't merely optimize a drug against a familiar target; its platform helped identify TNIK as a disease target before designing the molecule that eventually became rentosertib. A successful Phase III study would therefore say something more consequential than another ten-month medicinal-chemistry program.
It still wouldn't settle the question by itself. One successful drug can be an important proof point without establishing that AI-originated candidates are systematically more likely to work. Over the next several years, the more revealing number will be attrition: whether cohorts of AI-enabled drugs survive Phase II and Phase III at higher rates than comparable conventional programs, after accounting for disease, modality and the maturity of the underlying target.
Anthropic points toward another result worth watching. If agents can repeatedly surface genuinely novel biological systems or disease mechanisms that subsequently survive prospective laboratory testing, the improvement will have moved from searching chemical space into choosing what biology deserves to be acted upon. Cellular foundation models, perturbation datasets, genomic agents and other systems are all aimed somewhere in that direction, although it is precisely where the information is messiest and experimental validation remains indispensable.
For now, several clocks are moving at once. Schrödinger considered billions of compounds and reached a candidate in about ten months; Recursion moved new biological insight into IND-enabling work in less than 18; Anthropic's agents searched hundreds of thousands of proteins and surfaced a previously uncharacterized biological system in 21 hours; Insilico turned an AI-identified target into a drug that has now reached Phase III. None of those results establishes that AI-originated medicines are more likely to work, but the searchable part of biology is expanding well beyond molecules.
Rentosertib will spend the next 52 weeks helping answer the harder question. Claude needed 21 hours to find something unusual in a genome; unfortunately, at least for now, human lungs still need a year to tell researchers whether a drug works.