Home / Blog

AI and drug development

The biotech company with one employee and eleven laboratories

When Anthropic bought Mithril Medicine in July 2026, the company had one full-time employee and experiments running at 11 outside labs. Its technical report shows Claude speeding up the reading and email between experiments while the biology kept its own pace. That changes how many drug programs one scientist can run more than how fast any one of them finishes.

Main takeaways
  • One employee is a payroll count. Mithril bought its experiments from 11 outside labs and paid eight part-time consultants and advisers. Claude did the reading and drafting between them.
  • The antibody roughly doubled forelimb force in a small pilot. The better of two versions gave 1.95 times the force of untreated mice and closed 39.5% of the gap to healthy mice. Counted by litter, P was 0.057, the smallest value seven litters allow.
  • The labs set the pace. Claude’s first answers took a median of 96 seconds and the labs’ first replies 8.5 hours. Faster coordination shortens each program a little and raises the number one person can run.
  • Errors survived in the details. At least three of Claude’s errors reached a lab in emails that people had edited. The authors’ cases suggest that their editing checked decisions more reliably than details and premises.
  • Founding a company is a mixed hiring signal. In a field experiment, former founders got 43% fewer callbacks for software jobs. Workers who arrive through startup acquisitions leave more often than regular hires.

In July 2026 Anthropic bought Mithril Medicine, a drug company with one full-time employee. Its founder, Cody Tranbarger, had used Claude to organize contracts and experiments across 11 outside laboratories, according to Anthropic’s announcement on 27 October. Mithril works on two inherited forms of muscular dystrophy. In less than 18 months it took a treatment for one of them from an idea to an antibody that made sick mice stronger. Tranbarger now works at Anthropic, which says a drug could be ready for a first human trial as soon as late 2028. Anthropic labels every date after 2026 a target.

Anthropic calls the arrangement Pipeline-in-a-Person, one scientist working with Claude on several drug programs at once. Tranbarger and the Mithril team dated their technical report the day before the announcement. It has not been peer reviewed, and it is unusually frank. It reports the mouse data together with their weak points. It also times Claude’s answers against the labs’ replies and lists the errors Claude made. Read closely, the report supports a narrower claim than the headline. What Claude shrank at Mithril was the work between experiments, and that work sets how many programs one person can run.

Who did the work

A company with one employee can still buy the work of many people. Mithril bought most of its science from contract research organizations, or CROs, companies that run experiments for a fee. One CRO made Mithril’s antibodies and measured how tightly they bind to their targets. Another tested them on cells, and others ran the work in mice. Large drug companies buy research the same way, so the arrangement itself was ordinary.

Mithril also had a team. Anthropic says the rest of the work was done “by part-time consultants and third-party labs.” The report’s acknowledgments name eight team members. Its statement of competing interests says that members of the team “have been paid consultants or advisers to the company.” Their roles ranged from antibody engineering to translational planning, the work of preparing a drug for its first human trial. Figure 1 counts the people and labs the founder worked with.

A company of oneHow many people and labs did the founder work with?

5
1full-time employee 8part-time consultants and advisers 11outside labs Claude drafting in between
Everyone the founder of Mithril worked with The founder sits at the centre. Around him are eight paid part-time consultants and advisers, and around them 11 outside laboratories, five of which are labelled with their role in the 113-antibody screen. Claude circles the founder, reading and drafting. ClaudeProduction, bindingCell testsDevelopabilityMouse blood levelsMouse efficacyFounder

One full-time employee, eight paid part-time consultants and advisers, and 11 outside labs, with Claude reading and drafting between them.

Figure 1. Set a guess and press Reveal, or step through the layers. Point at or tap a circle for its role. Counts come from Anthropic’s announcement and the acknowledgments of the Mithril report. The five outlined labs are the CROs that the report lines up for its 113-antibody screen. The report does not give the other six labs’ roles, and the layout is a drawing.

Headcount describes how a company is organized. It says little about how much labour or money went in. Ten scientists on the payroll and ten scientists at a CRO all get paid, and only the first ten count as employees. Mithril has not published what it spent, so nobody outside the company can say whether its antibody cost less than it would have at a conventional biotech. The record does show how the work was split. The CROs ran the experiments. The founder and his advisers chose what to run, and Claude did the reading and drafting that connected the two.

Why companies hire anyone at all

Economists have asked for almost a century why firms employ people when they could buy every service on the open market. In 1937 Ronald Coase gave the standard answer in The Nature of the Firm. Using the market costs something beyond the price, he argued. A buyer has to find out what the relevant prices are, then negotiate and conclude a separate contract for each exchange. These are now called transaction costs. A firm hires staff when directing them is cheaper than paying those costs again for every job, and it buys from outside when the market is cheaper.

Claude lowers one part of those costs, the reading and writing around each job. In the week the report measured, a CRO delivered a 341-page report on 113 antibodies. Claude read every page and returned a summary of all 113 in 7 minutes and 53 seconds, with a proposed next step for each. Turning the first idea for a new drug program into a full plan took two to three days with Claude, and Tranbarger estimates it would have taken him one to two months without. That second figure is a guess about a path he did not take, so it carries less weight than the timed ones.

Coase’s logic also predicts which work stays inside a firm. A task that is hard to describe in a contract costs more to buy than to do, so the company keeps it. At Mithril that task was judgment. Claude had recommended measuring how tightly all 113 antibodies bind, including 24 that shared three parts with poor production results. On day 7 of the measured week, a team member challenged the plan. Claude defended it within 27 seconds, arguing that the data would show whether those parts deserved more work. The team member overruled it and set the 24 aside. In the report’s words, “one ruling on three arms decided 24 constructs.”

Coase’s trade-off, with invented numbersWhat happens as managing an outside lab gets cheaper?

Weeks of staff timeMinutes

Done inside

Choose what to test next

Bought from outside labs

Run the mouse study
Make the antibodies
Check they can be made as drugs
Measure how tightly they bind
Test them on cells
People needed inside 1

When managing an outside job takes minutes, one person can buy every experiment. Choosing what to test next stays inside at every setting.

Figure 2. Drag the slider. Each experimental job moves outside once managing it costs less than doing it in-house, and the count of people inside falls with it. The thresholds and head counts are invented to show Coase’s logic, and none of them are Mithril’s.

That kind of decision is hard to hand to a CRO, because the CRO’s job is to carry it out. As managing an outside job gets cheaper, more experiments can move out of the company, and the job of choosing what to test stays where it was.

An antibody that works like a clamp

Every muscle fibre is wrapped in a thin sheet of protein called the basement membrane, which holds the fibre in place as it contracts. The sheet is anchored to the fibre by a bond between two proteins. Laminin-211 sits in the sheet. Dystroglycan sits in the fibre’s outer membrane and has two parts. α-Dystroglycan, on the outside, carries a sugar chain that laminin grips. β-Dystroglycan crosses the membrane and connects to the scaffolding inside the fibre.

Mithril works on two diseases in which this bond fails. Children with LAMA2-related congenital muscular dystrophy, or LAMA2-CMD, inherit two broken copies of LAMA2, the gene for laminin-α2, one of laminin-211’s three chains. Their muscle makes a stand-in, laminin-411, which grips α-dystroglycan only weakly. In α-dystroglycanopathy, laminin is normal but α-dystroglycan lacks the sugar chain that laminin grips. Anthropic estimates that LAMA2-CMD affects 4,000 to 5,000 people in the US and Europe. The report cites a London group of 42 children with no laminin-α2 at all, in which one child learned to walk independently. Half were on a ventilator at night by age 11.7, and seven had died by age 17.

Bispecific antibody

An antibody is a Y-shaped protein whose arms bind a target. In an ordinary antibody every arm binds the same target. A bispecific antibody is engineered so that its arms bind two different targets. Most antibody drugs block a signal or mark a cell for destruction. Mithril’s antibody is meant to hold two proteins together, the way a clamp holds two boards.

Mithril’s antibodies grab β-dystroglycan with one arm and the stand-in laminin with the other. For α-dystroglycanopathy the same β-dystroglycan arm would pair with an arm for normal laminin. The approach has a history. In 2020 scientists at Sanofi reported a bispecific antibody that linked laminin to β-dystroglycan and improved muscle function in mice with an α-dystroglycanopathy. Engineered linker proteins had also helped mice with LAMA2-CMD. One of them, mini-agrin, raised median survival from 15.5 to 50 weeks when the mice made it themselves from an added gene. Mithril’s first two antibodies were built from published antibody sequences. Its own step was to aim the second arm at laminin-411 and test the result in LAMA2-CMD mice.

A simplified drawingWhat keeps a muscle fibre attached to its sheet?

The anchor between a muscle fibre and its basement membrane A protein sheet runs along the top and a muscle fibre along the bottom. Laminin hangs from the sheet. Dystroglycan sits in the fibre membrane, with alpha-dystroglycan carrying a sugar chain. In healthy muscle laminin-211 grips the sugar chain. In LAMA2-CMD the stand-in laminin-411 grips weakly. With the antibody, one arm holds laminin-411 and the other holds beta-dystroglycan. Protein sheetbasement membraneLaminin-411the stand-inα-Dystroglycanwith its sugar chainβ-Dystroglycanspans the membraneMuscle fibreAntibody

With the antibody. One arm holds the stand-in laminin and the other holds β-dystroglycan, which crosses the membrane, so the sheet stays attached.

Figure 3. Pick a muscle, then press Flex. The drawing follows the report’s description of the two laminins and dystroglycan, simplified and not to scale. The antibody shown has one arm for each target, like the 1+1 format described below. The 2+2 format has two arms for each.

Twice the force, in eleven mice

The test used dyW mice, which lack laminin-α2 and live about 12 weeks without treatment. The antibody groups were injected twice a week from birth for six weeks. Mithril tested two formats. The 1+1 version had one binding site for each target, and the 2+2 version had two of each. Both were given at the same number of molecules per injection, so the 2+2 delivered twice as many binding sites. The main measurement was the force of the triceps, a forelimb muscle, when its nerve was stimulated. These mice also lose the use of their hind legs through a nerve problem that rarely affects patients, so the forelimbs show the effect on muscle most clearly.

Mice given the 2+2 reached 1.95 times the peak force of untreated mice with the disease. The 95% confidence interval ran from 1.51 to 2.51. A 95% confidence interval is a range computed so that, across many repeats of a study, 95% of such ranges would contain the true value. This one excludes 1, which would mean no effect. Grip strength per gram of body weight rose by a similar factor, 1.94. The 1+1 version showed no detectable effect, at 1.08 times. Mini-agrin served as the positive control, a treatment already known to work, included to show that the test can detect an effect. Delivered once at birth by a gene-therapy virus, it reached 1.31 times.

The treated mice were still much weaker than healthy ones. The gain closed 39.5% of the gap between untreated and healthy mice, and treated mice still weighed about half as much as healthy mice. Figure 4 puts the three groups on one scale.

The report’s Figure 11B, redrawnWhere would healthy mice sit on this scale?

Peak triceps force of treated dyW mice relative to untreated dyW mice On a log scale, the 2+2 bispecific group has 1.95 times the untreated force, 95% CI 1.51 to 2.51, with 7 against 4 mice. Mini-agrin has 1.31, CI 1.00 to 1.72. The 1+1 bispecific has 1.08, CI 0.67 to 1.76. Healthy mice sit at about 3.4 times, and the 2+2 closed 39.5% of the gap. 0.5×1×2×3×4×no differencePeak force against untreated mice (log scale)2+2 bispecifictwo sites per target · 7 vs 4 mice1.95Mini-agrin gene therapypositive control · 4 vs 4 mice1.311+1 bispecificone site per target · 5 vs 4 mice1.08healthy, about 3.4×Gap to healthy closed39.5%by the 2+2, on a linear scale from untreated to healthy

The 2+2 group reached 1.95 times the untreated force. Healthy mice sit near 3.4 times, so the treatment closed 39.5% of the gap.

Figure 4. Tap the scale to guess where healthy mice sit, then see the answer. Squares are ratios of mean peak triceps force at week six against untreated dyW mice, and lines are 95% confidence intervals, on a log scale. The values are from Figure 11B of the Mithril report. The healthy level, about 3.4 times, is my calculation from the report’s statement that the 2+2 gain equals 39.5% of the deficit.

The confidence interval treats each mouse as an independent data point, and the report questions that choice. Each litter received a single treatment. Pups from one litter share a mother and grow up together, so their results tend to resemble each other. A P value is the probability of seeing a difference at least this large if the treatment did nothing. Counting the 11 mice as independent gives P = 0.0001. Using the average of each of the seven litters gives P = 0.057. Statisticians call the first approach pseudoreplication, treating repeated measurements of one unit as if they came from separate units.

With seven litters, 0.057 is the smallest P value the test can give. There are 35 ways to choose which four of seven ranked litters received the antibody. In the actual result all four treated litters ranked above all three untreated ones. Only one other arrangement is as lopsided, its mirror image with the treated litters at the bottom, so P = 2/35. Figure 5 lets you shuffle the labels and see how rarely chance produces a split that clean.

Seven litters, ranked by forceHow often does chance give a split this clean?

The 35 ways to place four treated litters among seven ranked litters Seven litters are ranked from weakest to strongest. In the real result the four antibody litters are the four strongest. A grid shows all 35 arrangements of four treated litters among seven, with the two most lopsided outlined. weakeststrongestantibody litteruntreated litter

Of the 35 ways to place four treated litters among seven, two are as lopsided as the real result, so P = 2/35 = 0.057.

Figure 5. Press Shuffle to relabel the litters at random, as if the antibody did nothing. Each small row in the grid is one of the 35 ways to place four treated litters among seven, and the two outlined in red are as lopsided as the real result. The real order, all four treated litters strongest, follows from the report’s P = 0.057, the minimum for seven litters. The litter averages themselves are not published, so the drawing shows only their order.
What the pilot cannot show yet

The groups were enrolled one after another, without randomization. The untreated mice also received buffer injections on a different schedule, weekly from day 7. The 2+2 mice were tested last, 126 and 161 days after the two sessions of untreated mice, and the colony moved to a new facility in between. Losses before week six left 7 of 9 treated mice and 4 of 8 untreated mice for the force test. Anthropic’s summary counts one death among eight treated mice against four among eight untreated mice. The report lists the untreated losses as three mice found dead and one removed early for a tissue test. It warns that “groups may differ by survival as well as by treatment.” Six treated groups and at least 11 measurements were examined, so a chance finding is possible. Safety was not tested.

The report calls the study a pilot and asks for a blinded replication with all groups run at the same time. The litter question is also a lesson about fast analysis. Counting mice gives a clean figure and a tiny P value, and the problem only shows when someone asks what the independent unit was.

Eighteen months, counted from where

Eighteen months from an idea to results in mice sounds fast. How fast depends on where the clock started and where it stopped, and a fair comparison needs another program measured the same way.

A program in obesity research offers one. In July 2022 scientists at Alnylam Pharmaceuticals reported a finding from the exomes, the protein-coding parts of the genome, of 362,679 people. Those carrying rare variants that disable the gene INHBE tended to have a lower waist-to-hip ratio for their body mass index, a sign of less fat around the abdomen. A variant that switches off a gene acts as a natural experiment, because it shows what lowering that gene’s activity does in people before anyone makes a drug. On 23 December 2024, about 29 months after the paper, Arrowhead Pharmaceuticals dosed the first participants in a trial of ARO-INHBE, a drug that silences the gene by RNA interference. Wave Life Sciences started a trial of its own INHBE-silencing drug on 6 February 2025.

The INHBE programs went further, into trials in people, and they started from a direct measurement in people of what lowering the gene does. Mithril’s starting evidence came partly from patients too. People whose LAMA2 variants damage only the end of laminin that links to other laminins usually walk, while variants in the anchoring end can cause severe disease from birth. That pattern suggests anchoring matters most, and anchoring is what the antibody restores. Mithril then built on published antibodies and earlier mouse work, and stopped at a pilot in mice. Claims that AI has compressed years of drug development into months raise the same question each time, which stretch of the path got shorter and compared with what.

Two clocks, two finish linesWhich program was faster?

Months from start to milestone for Mithril and two INHBE programs Mithril took under 18 months from an idea to pilot results in mice. Arrowhead dosed its first participant 29 months after the July 2022 INHBE paper, and Wave started its trial at 30 months. 0612182430Months from the starting pointMithrilfrom an ideaPilot results in mice, under 18 monthsArrowheadfrom a 2022 paperFirst person dosed, 29 monthsWavefrom the same paperTrial started, 30 months

Mithril stopped at a pilot in mice in under 18 months. The INHBE programs reached trials in people at 29 and 30 months, starting from a genetics paper.

Figure 6. Press Start to run the clock. Months from each program’s starting point. Mithril’s start date is not public, so its bar shows only the reported length, under 18 months. The INHBE bars run from the 27 July 2022 paper to Arrowhead’s first dosing on 23 December 2024 and Wave’s trial start on 6 February 2025.

Fast answers, slow labs

The report timed the week in which Mithril planned its largest screen so far, 113 antibodies across five CROs. Each antibody joined one of 14 arms that bind laminin with one of 8 arms that bind β-dystroglycan, plus one control. Claude’s first answer to a request came back in a median of 96 seconds. Its full answer, a plan or a set of drafted emails, took a median of 10 minutes. The CROs’ first replies took a median of 8.5 hours, a figure that includes nights and a weekend as well as a 12-hour time difference. The report’s conclusion is that “the pace of this work was therefore set by the replies of the CROs and by our decisions.”

One clock, from seconds to monthsWhere does the time go?

Times from Mithril's measured week and program on a logarithmic clock Claude's first answer took a median of 96 seconds, reading a 341-page report took 7 minutes 53 seconds, and its full answer a median of 10 minutes. A lab's first reply took a median of 8.5 hours. Eleven open questions were addressed in four working days, and 89 antibodies shipped within a week. Mouse dosing took six weeks, and the program took under 18 months from idea to mouse results. 10 s1 min1 hour1 day1 week1 month1 year96 secondsClaude’s first answer (median)7 min 53 sClaude reads a 341-page report10 minutesClaude’s full answer (median)8.5 hoursA lab’s first reply (median)Four working daysEleven open questions addressedUnder a week89 antibodies shipped to the labsSix weeksDosing in the mouse studyUnder 18 monthsFrom an idea to results in mice

Claude’s steps took seconds to minutes. The labs’ replies took hours, and the biology took weeks.

Figure 7. Choose a group, or point at a dot. The scale is logarithmic, giving equal space to each tenfold step in time. Claude’s and the labs’ times are medians or single events from the week the report measured. The dosing period and the 18-month figure come from the report and Anthropic’s announcement.

Computer scientists have a name for this limit. In 1967 Gene Amdahl argued that speeding up one part of a job shortens the whole job only so far, because the rest still takes its full time.

Amdahl’s law

If a fraction p of a job can be made s times faster, the whole job speeds up by a factor of 1 / [(1 − p) + p/s]. A ten-month program might include four months of coordination on its critical path, the chain of steps that each wait for the one before. If those four months shrink by 75%, they become one month and the program takes seven. The coordination became four times faster, and the program became 30% shorter.

Anthropic is betting on a different number. In 1961 John Little proved a rule that holds for any system with work in progress. The average number of jobs in progress equals the rate at which jobs are finished multiplied by the average time each one takes. Rearranged, the number of programs a scientist finishes per year equals the number they keep going at once, divided by how long each one takes. Biology sets how long a mouse study takes. Coordination limits how many programs one person can keep going. Pipeline-in-a-Person is an attempt to raise the second number while the first stays put.

Two rules, with invented numbersFaster programs, or more of them?

Amdahl’s law, one ten-month program

40%
4×
A ten-month program before and after faster coordination A ten-month program has six months of biology and four of coordination. With coordination four times faster it takes seven months, 30 percent shorter. biology 6.0 mo7.0 months, 30% shorterbefore, 10 months

Little’s law, programs per person

3
Program 1
Program 2
Program 3

The program takes 7.0 months, 30% shorter than ten. Even instant coordination would leave 6.0 months. Running 3 at once, one person finishes about 5.1 programs a year.

Figure 8. Move the sliders. The top bar applies Amdahl’s law to a ten-month program, with biology in blue and coordination in gold. The lanes apply Little’s law, in which the programs finished per year equal the number running at once times 12, divided by the months each takes. All numbers are invented to show the two rules.

More programs per person has a cost that the formula leaves out. Each program brings its own data and its own decisions. If one person reviews more decisions in the same hours, each decision gets less attention, and more errors can reach the labs. The report shows where Claude’s errors went.

Where the errors went

In the week the report measured, Claude sent no emails. It saved drafts in a team member’s mailbox, and a person edited and sent each one. Over the measured week Claude drafted 17 emails and the team sent 12. For ten of the sent emails the report kept both versions, which shrank from 4,938 words as drafted to 3,409 as sent, a cut of nearly a third. With its drafts Claude also listed 41 points where it had found no written instruction from the team, such as how much to put in each tube. A person could then confirm or change each one.

The authors examined six of Claude’s errors in detail. Claude itself found five of them, two before the affected email went out and one only after a CRO asked a question. At least three errors reached a CRO in emails the team had edited and sent. One, a request for less material per tube than the lab had asked for, was fixed within a day, before any tube was filled. Two were statements in Claude’s review of the production data that went unnoticed for more than four days. One of those is still unsettled, and it concerns the evidence on two of the three arms that were set aside on day 7.

The most useful sentence in the report is the authors’ reading of these cases. Their editing, they write, “tests the decisions in an email more reliably than the details and premises that it carries.” A person reading a draft asks whether the plan is right. The numbers and assumptions under the plan get less attention, and fluent writing makes them look as finished as everything else. The authors add a caveat of their own. Claude compiled the section on working with Claude from its own threads and notes, which, they write, “carries a risk of bias.” Reporting the errors was their attempt to limit it.

An invented email, with three slipsCan you check it in 25 seconds?

Your notes, all correct
  • After Tuesday’s decision, 89 antibodies go to the cell tests.
  • Lab B asked for 50 µL in each tube.
  • The three set-aside arms were held back for poor production results.
Draft by Claude · to Lab B · Plan for the cell tests

Thank you for the plate map, which matches the layout we asked for. We will ship all 113 antibodies for both cell tests on Friday.Wrong decision. After Tuesday the plan is 89 antibodies. Please include the reference antibody on every plate. Each tube will hold 20 µL, as you requested.Wrong detail. The lab asked for 50 µL. We set three arms aside because they bound laminin too weakly.Wrong premise. They were held back for poor production results. I will confirm the shipping address by Thursday.

The email carries one wrong decision, one wrong detail and one wrong premise. In Mithril’s week, the team’s editing caught wrong decisions more reliably than wrong details and premises.

Figure 9. Start the clock, then tap every sentence that disagrees with your notes. The email and notes were written for this post. Two of the slips copy kinds the report describes, a wrong amount per tube and a wrong statement about the data behind a decision. The 25-second limit is arbitrary.

Buying the company to get the founder

Anthropic could have offered Tranbarger a job. Buying Mithril brought more with him. Anthropic now owns the antibody program, according to the report, and by October that program held more than 100 new bispecific antibodies built from Mithril’s own binding arms. The purchase also brought a working method that Anthropic plans to test on new programs. The terms of the deal have not been disclosed, so any claim about retention bonuses or shares that vest over time would be a guess.

Research on acquisitions suggests why the circumstances matter. J. Daniel Kim compared workers who joined a firm when it bought their startup with workers hired the ordinary way, using population-level data from the United States. Acquired workers left at significantly higher rates. The gap was smaller when the founding team stayed intact and when the buyer kept the startup structurally separate. Mithril’s founding team was one person, and he still runs both of its original programs.

Whether founding a company helps someone get hired is also less clear than it sounds. Tristan Botelho and Melody Chang sent applications to 2,400 software engineering jobs in the United States, varying only the applicant’s history as a founder. Former founders got 43% fewer callbacks than applicants who had never founded a company, a gap driven by older hiring firms. Founders whose companies had succeeded fared worst, with 33% fewer callbacks than founders whose companies had failed. The authors trace the penalty to employers’ doubts about whether a former founder would fit in and stay.

A field experiment with 2,400 applicationsWhich applicant gets the most callbacks?

Former founders got 43% fewer callbacks than applicants who never founded a company
Founders who succeeded got 33% fewer than founders who failed

Employers called back applicants who had never founded a company most often. Founders whose startups had succeeded came last.

Figure 10. Tap the three applicants in the order you think employers called them back, most callbacks first. The applications were identical apart from founder history. The two ratios are from the abstract of Botelho and Chang (2023). Together the two ratios fix the order of all three groups, because failed founders must then get at most about 85% of the non-founders’ callbacks.

The two studies fit together. A founder brings evidence of what they can do, and employers discount it because they doubt the founder will stay. Buying the company answers part of that doubt, since the buyer keeps the work even if the person later leaves. Cheap, fluent writing also makes polished descriptions of past work less informative. A report like Mithril’s is harder to fake than a pitch, because it counts its own errors and reports the analysis that weakens its main result.

What changes for the next founder

For someone starting a biotech, the Mithril record moves the scarce skill. Writing a plan has become cheap. On day 4 of the measured week, Claude turned a 341-page data package into a proposed next step for each of 113 antibodies in under eight minutes. The slow part was a person deciding which proposals to accept, and one of those decisions removed 24 antibodies from the screen.

The model also carries a risk that the report makes visible. A company that buys every experiment depends on its labs to run protocols well and to report problems, and its founder sees the work only through files and emails. In Mithril’s week, the errors that lasted longest were statements inside those files.

The next test runs on Anthropic’s timetable. Mithril aims to choose formal drug candidates for both diseases in 2027 and to start a LAMA2-CMD trial as soon as late 2028. Tranbarger is also preparing several more programs, all for childhood diseases with no approved therapy.

Further reading

AI for science drug development economics of the firm statistics