Eli Lilly announced on August 26 that a real-world analysis of adults over 55 with obesity found sustained Zepbound use associated with lower monthly healthcare costs than matched untreated adults. The release leads with an estimated difference of up to $607 per patient per month, roughly 38% lower costs. That number is in the paper. It is also the output of an analysis the paper itself demotes.

The primary analysis reports $319 per patient per month, a 25.4% reduction, with a confidence interval running from $94 to $544. Same cohort, same window, and a spread wide enough that at the top end the offset comfortably exceeds what Medicare pays for the drug while at the bottom end it covers less than half. Which figure you carry into a conversation decides what this study says.

There are 3 claims in play here, and keeping them apart is most of the work. Lilly's release implies the drug may pay for itself. The paper says something narrower, that it may offset a meaningful portion of its cost among people who stay on it. And the argument neither one makes, which is the strongest of the three, is that a treatment does not have to pay for itself to be worth covering.

The paper deserves better than the reflexive dismissal it will get from clinicians who notice that 3 of the 6 authors are Lilly employees and shareholders while the other 3 work for firms Lilly paid to run the analysis. The underlying observation is mechanistically sensible. But Lilly has chosen to fight on the weaker of the grounds available to it, and the stronger case for keeping GLP-1 coverage in Medicare was made 18 months ago by academics with no stake in the answer.

This was a retrospective, propensity-matched cohort study using the Komodo Research Dataset, a de-identified US claims database covering more than 330 million lives, over a study period running from November 2022 through September 2025. Adults over 55 with obesity, or with overweight plus at least one weight-related comorbidity, and crucially without type 2 diabetes, who initiated tirzepatide for weight management between November 2023 and September 2025 were matched 1:1 to controls who never received any incretin-based therapy. The over-55 group of 15,843 pairs was a pre-specified subgroup carved out of a larger matched population of 171,299 pairs aged 18 and up.

The exclusion of diabetes is a point in the study's favor rather than against it, since it isolates the obesity signal instead of letting glycemic control and diabetes-related admissions carry the result. Outcomes were assessed with a difference-in-differences framework across 3 follow-up bands, 3 to 6 months, 6 to 12 months, and 12 to 18 months, with the first 3 months dropped to avoid the dose-titration period. Every cost figure excludes the price of tirzepatide itself, which the authors state plainly.

Two censoring schemes were run, and the relationship between them is the most interesting thing in the paper. Pairwise censoring, which censors a patient when their matched partner is censored, was pre-specified as the primary approach. Inverse probability of censoring weighting was the pre-specified sensitivity analysis. Then, in the authors' own words, post hoc inspection revealed how few patients remained in the 12 to 18 month band, which raised the likelihood of informative censoring between arms, and on that basis the IPCW results were promoted to primary while pairwise was retained as supporting. The swap moved toward the more conservative estimate, which is to the authors' credit. Lilly's press release then leads with the number from the analysis that got demoted.

How few is few? Of 15,843 pairs at baseline, the pairwise scheme leaves 793 per arm in the 12 to 18 month band. Under individual censoring the imbalance is starker: 1,181 treated patients against 3,004 controls, which is precisely the informative censoring the authors flag. The paper's limitations section concedes that fewer than 10% of individuals remained uncensored in that band, reducing precision and generalizability and preventing any subgroup analysis. The $607 that will be quoted in coverage debates for the next year rests on 793 matched pairs.

It cannot fully separate the drug from the person taking it. The comparator is never-treated, not treated-and-discontinued, and the authors concede in their limitations that residual confounding from socioeconomic status, lifestyle behaviors, and health literacy cannot be excluded. Two details sharpen this. Patients initiated during a window when Medicare Part D was statutorily barred from covering Zepbound for weight management, so persistence for a full year generally required a pharmacy benefit willing to pay, though the obstructive sleep apnea indication opened a narrower legitimate route in the same period. And medications obtained through cash-pay, compounding, or direct manufacturer supply were not captured at all, meaning some genuinely treated people sit in the untreated control arm. That second bias runs the opposite way and would understate the true difference, which is worth saying out loud. In the other direction, the treated cohort was slightly sicker at baseline, with more weight-related comorbidities and more sleep apnea, which also cuts against a simple healthy-adherer story.

It cannot tell you whether treatment saves money. The drug's own cost is excluded from every estimate, and to the authors' credit they say so plainly and explain why: rebates paid directly to payers are not captured in claims, so including a list-price figure would distort more than it clarifies. What that leaves is an offset analysis, which asks how much of the price downstream care hands back, not whether the whole transaction nets positive.

It cannot speak to a therapy people take indefinitely. 18 months is a reasonable first look and an unreasonable basis for a coverage decision about a disease treated for decades, particularly when the authors state that findings apply only to people who stay on treatment and cannot be extrapolated past discontinuation. They also declined to run an intention-to-treat analysis, citing the 2024 supply shortage that affected persistence for the 36% of the cohort indexed that year. That is a defensible call. It also means the most policy-relevant estimand, what happens to costs across everyone who starts the drug rather than only those who stay on it, is absent.

That estimand is not entirely absent from the literature, and the closest attempt at it points the other way. In a January 2026 working paper, Wing and colleagues used commercial claims from 2017 through 2024 in a stacked difference-in-differences design that keeps discontinuers in the denominator. They found no offset. Every dollar of GLP-1 spending was associated with roughly 22 cents in additional non-GLP-1 medical spending. That cohort is younger, commercially insured, and drawn from an era when most GLP-1 initiation was for diabetes, and the paper has not been peer reviewed. It is still the nearest thing available to the analysis Lilly's authors declined to run, and it did not find what the press release describes.

Now the part that should give any reader pause. Lilly attributes the savings to fewer hospitalizations and emergency visits, and in the IPCW analysis that holds well, with incidence rate ratios of 0.86, 0.76, and 0.69 across the 3 bands, all statistically significant. But in the pairwise analysis, the one that generates the $607, the inpatient and ED rate ratio at 12 to 18 months is 0.99, with a confidence interval from 0.60 to 1.64 and a p value of 0.980. That interval is not evidence the mechanism vanished. It is compatible with a 40% reduction and with a modest increase, which is what 793 pairs buys you. The problem is narrower and harder to argue with. The analysis producing the largest cost saving is too small to demonstrate the utilization change that is supposed to explain it.

The paper's own conclusion is carefully worded: tirzepatide coverage may partially offset the drug cost by 12 months and may result in cost savings by 18 months, among those who stay on therapy. The benchmark is the Medicare GLP-1 Bridge, under which Novo Nordisk and Lilly supply covered GLP-1s at a negotiated net price of $245 per 30-day supply. The beneficiary pays a flat $50 at the pharmacy regardless of dose or income, and the remaining $195 is what the federal government carries.

So $195 is not the treatment's cost. It is the taxpayer's share of it, which is a legitimate number if the question is what the Bridge costs CMS, and it is the smaller of two defensible denominators. The more important omission is the range. The paper's abstract gives the offset as $319 with bounds of $94 to $544, and states the $94 figure explicitly in its conclusion. The press release drops the range entirely and replaces it with a larger point estimate from the supporting analysis. At the lower bound the offset covers less than half of Medicare's $195 share, let alone the full $245. None of this is misconduct. It is the ordinary distance between a paper and its press release, and it is the reason to read the abstract rather than the announcement.

In March 2025, Hwang and colleagues at the University of Chicago published a lifetime cost-effectiveness analysis in JAMA Health Forum using a validated microsimulation model applied to NHANES data. Tirzepatide came out with the largest incremental QALY gain of any anti-obesity medication studied, 0.35, and an incremental cost-effectiveness ratio of $197,023 per QALY, which is roughly twice the conventional threshold. To reach $100,000 per QALY, they calculated that tirzepatide's price would need to fall by 30.5%, from about $6,236 per year to about $4,334. Both of those are net prices, not list, which is what makes the next paragraph a fair comparison rather than a rhetorical one.

The Bridge net price is $245 a month. That is $2,940 a year, roughly 32% below the threshold price Hwang's group calculated, and it gets there without a claims database, without an assumption about hospital admissions, without a two-fold spread between models, and without an author list employed by the manufacturer. The policy argument that GLP-1 coverage in Medicare is defensible on value grounds was already available, already independent, and already stronger than the one being made this week.

The comparison is not clean, and I will not pretend otherwise. Hwang modeled a general adult population with a mean age of 48 rather than adults over 55, used a lifetime horizon rather than 12 months, and assumed a persistence pattern that real-world discontinuation rates do not match. Different population, different question, different uncertainty. But what matters here is the direction of travel, and the price cut has already done the work the offset study is straining to do.

The same group published a second paper that cuts the other way, and leaving it out would be the selective quotation this piece is complaining about. In April 2025, Hwang and colleagues modeled the 10-year fiscal impact of Medicare Part D coverage for GLP-1 receptor agonists and projected $65.9 billion in drug costs against $18.2 billion in health care savings, for $47.7 billion in net new federal spending. That analysis used pre-Bridge net prices, and a $2,940 annual price moves the larger of those 2 numbers substantially. It does not make it disappear. Covering obesity treatment in Medicare costs the program money. The defensible claim is that the money buys more health per dollar than a great deal of what the program already pays for, not that the drug pays for itself.

This is a competent study of a question that badly needed one, and the finding that treating obesity in older adults slows the growth of downstream spending is both mechanistically sensible and probably true. What I would not do is carry the $607 into a conversation with a payer or a policymaker. The paper's own primary analysis says $319, its confidence interval runs down to $94, and the model that generates the larger figure is too small to show the reduction in hospitalizations that is supposed to explain it. The person across the table will eventually read the paper. Quote the primary estimate, name the attrition, and the argument survives contact.

Three practical things follow for anyone seeing patients over 55. Check Bridge eligibility deliberately rather than assuming these patients are still uncovered, because the tiers are specific and the coverage is real. Document BMI and qualifying comorbidity carefully at the moment of initiation, since eligibility is anchored to the BMI recorded when GLP-1 therapy began. And raise the December 2027 expiration at the start of treatment rather than at the end of it, because a patient who understands from day one that this coverage has a defined horizon can plan around it, while a patient who finds out at the pharmacy counter cannot.

The case for covering obesity treatment in Medicare does not depend on the drug paying for itself. It depends on the drug being worth what it now costs, and at $245 a month, independent modeling published more than a year ago already said it is.

REFERENCES