The story goes like this. Artificial intelligence has learned to read the language of molecules, so the slowest and most expensive process in medicine is about to become fast. A drug that once took four years to find now takes eighteen months. The people running the world’s largest AI labs say most disease could be finished inside a decade. And if medicine starts handing you more than a year of extra life for every year you spend waiting, then nobody alive today has to die of anything in particular. That is the 2030 promise. The first half of it is already true.
The pill that arrived years early
Idiopathic pulmonary fibrosis scars the lungs until they stiffen and stop moving air properly. Nobody knows what starts it. Most people diagnosed with it die within a few years, and the two drugs licensed to treat it slow the decline rather than stopping it. It is exactly the sort of disease that defeats conventional drug hunting, because to find a treatment you first have to work out which of roughly twenty thousand human proteins is worth attacking.
In 2026 a tablet called rentosertib entered a final-stage human trial for it. The target it attacks, an enzyme called TNIK, was picked out of the field by a machine learning system called PandaOmics. The molecule that attacks it was drawn by a second system, Chemistry42. Neither the choice of target nor the shape of the drug began in a human idea.
The timings are the part that made the industry sit up. Insilico Medicine, the company behind it, says it went from starting the project to naming a preclinical candidate in about eighteen months, against an industry norm it puts at two and a half to four years, and that it did so having made and tested only 60 to 200 molecules. Target discovery to first human dose took under thirty months. In July 2026 the drug began a Phase III trial of 320 patients across 47 centres in China. No drug discovered by generative AI had ever got that far.
Finding a new medicine is like trying to pick one particular key out of a warehouse of billions, when you have never seen the lock. Chemists used to wander the aisles grabbing armfuls of keys and trying them. The AI looks at a description of the lock and draws a key. Insilico drew about a hundred keys instead of grabbing a hundred thousand, and one of them turned.
Rentosertib is not a lucky one-off. In 2024 a team from Boston Consulting Group published the first systematic look at how AI-discovered molecules behave once they reach people. They found that in Phase I, the first test in humans, those molecules succeeded at a rate of 80 to 90 per cent, against a historic industry range of roughly 40 to 65 per cent. Chemistry designed by machine is, on this evidence, unusually well behaved inside a human body.
Antibiotics tell a similar story, and a more urgent one, because drug-resistant bacteria kill somewhere near five million people a year and the pipeline for new ones has been thinning for decades. A group at MIT led by James Collins used generative models to invent compounds that do not exist in any catalogue. In one arm of the work the models produced roughly seven million candidate molecules built around a promising fragment; screening narrowed that to about a thousand, then to eighty worth making, of which chemical suppliers could actually synthesise two. One of those two, named NG1, cleared drug-resistant gonorrhoea in mice and did it by jamming a bacterial protein nothing else targets. A second arm, where the models were allowed to design freely, produced DN1, which cleared an MRSA skin infection in mice. In November 2026 Collins and his colleague Regina Barzilay were given the Pritzker Prize for this line of work.
Read the first two of those numbers and the conclusion writes itself. The men running the labs have certainly read them. Demis Hassabis, who shared a Nobel Prize for the protein-folding model that opened this whole field, went on American television in April 2025 and said the end of all disease was within reach. Dario Amodei of Anthropic has written that AI-enabled biology could compress fifty to a hundred years of progress into five or ten, and said in August 2026 that it should be possible to cure most human disease in that window.
Maybe within the next decade or so, I don’t see why not.Demis Hassabis, CBS 60 Minutes, 20 April 2025, on curing all disease
The third number nobody quotes
Go back to that same Boston Consulting Group paper and read one line further. AI-discovered molecules clear Phase I at 80 to 90 per cent. They clear Phase II at about 40 per cent, which the authors describe, carefully, as comparable to historic industry averages.
That single sentence is the whole argument of this article.
Phase I asks whether a molecule behaves itself: whether it dissolves, whether it reaches the tissue, whether it poisons anyone at a useful dose. That is a chemistry question, and chemistry is exactly what these models are good at. Phase II asks something different. It asks whether the thing the drug does to the body actually makes a sick person better. That is a biology question, and on that question the machines are currently no better than we were.
Which means the headline is not that AI halved drug development. It is that AI got very good at one stage and left the others where it found them. And the stages it left alone are where the years live.
Figure 1 — Illustrative schematic. Stage lengths are typical industry ranges, not a measurement of any single programme.
Rentosertib makes the point better than any diagram. Its Phase III trial measures how fast patients lose lung capacity over 52 weeks. There is no model, no cluster, no amount of compute that turns 52 weeks of human lung function into 20. You have to wait, and then you have to wait again while the data is cleaned and the regulator reads it. Insilico shaved perhaps two years off the front of that drug’s life. The back of it runs on the same calendar it always did.
The most expensive illustration of the gap is Isomorphic Labs, the Alphabet company Hassabis founded to turn protein prediction into medicines. It raised 2.1 billion dollars in May 2026 and by October was reportedly in talks at a valuation between 40 and 50 billion. As of May 2026 it had not given a single molecule to a single human being. Its first trials, promised for the end of 2025, were moved to the end of 2026. In November its president, Max Jaderberg, told a conference in London that the company was in its first innings and predicted that AI would one day design every medicine. Both things are true at once: the prediction may well be right, and there is still no patient.
The number behind “ten years and billions of dollars”
The figure everyone quotes comes from a 2014 study by the Tufts Center for the Study of Drug Development: 2.56 billion US dollars per approved drug, in 2013 money, over a process it described as often lasting more than a decade. Only about 1.4 billion of that is money actually spent. The rest is the return investors give up by having their capital tied up for ten years.
That composition matters for the promise. Roughly half the cost of a drug is the cost of time, and the single largest block of that time is the part AI has not touched.
Chemistry is a search. Biology is a guess
There is a clean reason the machines help enormously at one stage and barely at the next, and it has nothing to do with how clever the models are.
Designing a molecule is a search problem. The space of possible small molecules is unimaginably large, the rules governing it are physical and consistent, and there is a great deal of training data about what makes a molecule behave like a drug. Those are the conditions under which modern machine learning is overwhelming. Give it a well-posed search with good feedback and it will beat any human team that has ever existed.
Choosing what the molecule should attack is not a search problem. It is a question about a system nobody has fully characterised. Which protein, in which tissue, in which cell state, in which subset of patients, is doing the damage? The data that would answer it properly does not exist, because it would require running the experiment on living people. So the target is, in the end, an informed guess, and Phase II is where the guess gets marked.
Imagine a car making a terrible noise. AI is now a world-class mechanic: name any part and it will build you a perfect replacement overnight, beautifully machined, guaranteed to fit. What nobody can tell you, including the AI, is which part is making the noise. Phase II is opening the bonnet after you have fitted the beautiful new part, to find out whether the noise has stopped.
We can put a figure on how much the guess matters. In April 2024 a team led by Eric Minikel at the Broad Institute published an analysis in Nature of what separates drugs that get approved from drugs that fail. Only about one clinical programme in ten ever reaches approval. But programmes whose target is supported by human genetic evidence, meaning there is a real inherited variant in real people linking that protein to that disease, succeed about 2.6 times more often than programmes without it. The benefit is not uniform: it varies by disease area and by the stage of development.
Sit that finding next to the AI results and the shape of the problem becomes obvious. Machine learning is currently strongest where genetics matters least, and weakest where it matters most. We have automated the part of the job that was already tractable.
Now take that back to the longevity claim, because this is where the arithmetic gets uncomfortable.
Longevity escape velocity has a precise definition. Aubrey de Grey, who coined the actuarial version of the term, and David Gobel, credited with the phrase itself, both mean the same thing: the point at which medicine adds more than twelve months to your remaining life expectancy for every twelve months that pass. Below that line your horizon keeps retreating and you die on schedule. Above it, the horizon runs away from you faster than you can reach it. Ray Kurzweil put the crossing at 2029, and writing in The Economist in 2024 gave a range of 2029 to 2035. De Grey has given it a 50 per cent chance by the late 2030s. There is no fuzziness in the definition. The threshold is twelve.
So what rate are we actually running at? In 2018 Carolina Cardona and David Bishai published an analysis of 139 countries covering 1950 to 2009. Across the whole sample, the average gain in life expectancy fell from 9.7 years per decade in the 1950s to 1.9 years per decade in the 2000s. For the healthiest countries, those already above 71 years, the decadal gain fell from 4.8 years to 2.4. Progress has not been accelerating. It has been slowing for seventy years, through antibiotics, through statins, through the entire molecular biology revolution.
Nineteen-tenths of a year per decade is about 2.3 months per year. Escape velocity needs twelve. Move the slider below and see what that distance means.
How much life would medicine have to buy you?
Imagine someone with 32 years of life still ahead of them. Each year that passes costs them one year of that. Medicine hands some of it back. Set how many months it hands back annually, and watch what happens to the time they have left.
One honest caveat: the 1.9-years-per-decade figure measures life expectancy at birth, while escape velocity is defined on remaining life expectancy. They are not the same measure, and a dial is not a forecast. What the comparison is good for is the order of magnitude, and the order of magnitude is a factor of five.
The revolution is real. It’s just a different one
Here is the thing the sceptical account usually gets wrong, and it deserves as much emphasis as everything above. The interesting consequence of AI in medicine was never going to be speed. It is range.
For a century the practical definition of a medicine was a small molecule you could find, make and store, or a protein you could borrow from something living. That constraint is dissolving. Whole categories of treatment that were previously impossible to design are now being designed, and several of them have no conventional equivalent at all. Pick one and see what it is.
Two of those entries are worth sitting with, because they show where this actually goes.
In November 2026 doctors treated a patient whose E. coli infection had shrugged off every antibiotic and every phage they could find. A phage is a virus that eats bacteria, and the usual approach is to go looking in sewage for one that happens to fit. Instead a company called Radical Numerics used a model to redesign the gripping legs of a known phage so they would clamp onto this patient’s particular resistant strain. The DNA was synthesised, the viruses were grown and screened, and the FDA authorised the treatment. The company’s chief executive has said he hopes eventually to build a patient’s phage in days.
Hold the enthusiasm at the right level. That work has been posted as a technical report and has not been peer reviewed or published in a journal. It is one patient. Outside experts have been interested and cautious in roughly equal measure, and results will take months. But notice what it is not: it is not a faster version of something we already did. There was no previous way to do it at all.
The second is quieter. In October 2026 Anthropic disclosed that it had bought a biotech called Mithril Medicine which, before the acquisition, had exactly one full-time employee. Its founder used Claude to coordinate contracts, logistics and experiments across eleven external laboratories, and took an antibody treatment for a rare childhood muscular dystrophy from idea to promising results in mice in under eighteen months. Human trials, if the candidates hold up, are hoped for as early as late 2028.
One person. Eleven labs. A disease too rare to have ever justified a conventional programme. If that pattern holds, the thing AI changes is not the speed of the pipeline but the economics of entering it, and the diseases that benefit most are the ones that were never worth anyone’s time.
| Medicine | What it is | Furthest stage reached |
|---|---|---|
| Rentosertib | Pill for lung scarring; target and molecule both AI-selected | Phase III (320 patients, from July 2026) |
| NG1 / DN1 | Antibiotics invented from scratch for resistant gonorrhoea and MRSA | Mice |
| Redesigned T7 phage | Virus rebuilt to grip one patient’s resistant bacteria | One patient, FDA-authorised; not peer reviewed |
| OpenCRISPR-1 | Gene editor with every component AI-designed | Published in Nature; released as a research tool |
| Mithril bispecifics | Antibodies for a rare childhood muscular dystrophy | Mice; human trials hoped for late 2028 |
| RetroSOX / RetroKLF | Improved versions of the proteins that rewind cells to a younger state | Cells in a dish; not peer reviewed |
| Isomorphic Labs pipeline | Cancer and immunology drugs from the AlphaFold lineage | No human dosed as of May 2026 |
Figure 2 — Stages as reported by developers and journals. Mice are not people, and a technical report is not a peer-reviewed result.
Think of it like cooking. The promise was that a robot chef would get dinner on the table in ten minutes instead of two hours. What actually happened is that dinner still takes two hours, because the oven takes two hours, but the robot can now invent dishes out of ingredients nobody knew were edible. That is a smaller promise and a stranger one, and in the long run it is probably worth more.
So what should you expect by 2030, if you are a person rather than a slide in a pitch deck?
Expect the first approvals. Rentosertib’s trial reads out after 52 weeks of treatment, and if it works, an AI-discovered drug will be on pharmacy shelves before the decade turns. Expect two or three more behind it. Expect new antibiotics to enter human testing, which matters more than almost anything else on this list. Expect bespoke treatments, built for one patient at a time, to move from a single newspaper story to a small and carefully watched programme. Expect rare diseases that have never had a drug programme to get one, because a single person with a model and eleven contract labs can now afford to try.
Do not expect your remaining life expectancy to start growing faster than you are ageing. Nothing in the current evidence points at that, the one measured number we have has been moving the wrong way for seventy years, and the stage of drug development that would have to collapse for it to happen is the stage that has not moved at all. Derek Lowe, a working medicinal chemist who has spent two decades writing about these promises, put the objection plainly after the Hassabis interview: our understanding of disease is nowhere near adequate to end it in ten years, and it will still be inadequate at the end of those ten years.
That is not a small or disappointing future. A decade from now we may be able to make a medicine for a single person, for a disease that affects four hundred people worldwide, in a laboratory that fits in one room. Measured against what medicine could do in 2020, that is extraordinary. It just isn’t immortality, and the distance between the two is the distance between a chemistry problem we have largely solved and a biology problem we have barely started.
Buy me a coffee