Veterans who started semaglutide or tirzepatide had roughly a third the overdose risk of veterans who started insulin over the next year. Next to metformin, the same two drugs showed no advantage at all.

That split is the whole story. A target trial emulation ↗ published July 15 in the Journal of Clinical Psychiatry ran the same drugs against five different diabetes medicines, and whether the GLP-1 drugs looked protective depended entirely on which comparison drug sat on the other side.

What the study did

The authors pulled Veterans Health Administration records for patients who carried both type 2 diabetes and a diagnosis of opioid use disorder, and who started a diabetes drug between January 2020 and December 2024. A target trial emulation is an attempt to squeeze a randomized-trial structure out of ordinary medical records. You define who would have been eligible, when their treatment clock starts, and what counts as the outcome. Then you match the groups on background differences before comparing them. Here the treatment group was semaglutide and tirzepatide taken together, the outcome was any overdose (not only opioid overdoses) in the 12 months after starting, and each comparison group was matched separately on patient characteristics.

The verdict flips with the comparator

Against insulin, the GLP-1 drugs were tied to a 68 percent lower overdose risk (hazard ratio 0.32, 95 percent confidence interval 0.14 to 0.71, 630 patients). Against SGLT2 inhibitors, the class that includes empagliflozin, the reduction was 75 percent (HR 0.25, 0.09 to 0.68, 432 patients). Both were statistically clear.

Then the signal disappears. Against metformin, the first drug most people with type 2 diabetes ever take, the estimate was HR 0.75 (0.38 to 1.45, 1,016 patients), with a confidence interval wide enough to include no effect. Against sulfonylureas it was 0.82 (0.33 to 1.96). Against DPP-4 inhibitors the point estimate actually sat above 1.0 (HR 1.20, 0.52 to 2.78), meaning slightly more overdoses, though again the interval crosses no difference. Three comparators, three null results.

Why the metformin comparison is the one to watch

The comparators where the GLP-1 drugs won are the ones prescribed to sicker patients. Insulin tends to come later, when other drugs have failed, and the people on it are frailer and higher-risk at baseline. Matching narrows that gap but rarely closes it. Metformin is the closest thing to a healthy-user comparator in this list, the drug given early to people who are otherwise doing reasonably well, and it is exactly there that the overdose advantage vanishes. When a benefit shows up against the last-line drug and evaporates against the first-line one, the cleaner reading is that some of the effect belongs to who takes which drug, not to the drug itself.

The numbers are also small. The largest comparison held 1,016 patients and the smallest 432, and overdoses are rare events, so a handful either way moves a hazard ratio. This is a signal worth chasing, not a settled result.

Where it fits

GLP-1 drugs reached opioid use disorder through the reward system. Animal work has shown that activating the GLP-1 receptor blunts drug seeking, and the same logic has driven human observational hints on drinking ↗ and on opioid use in chronic pancreatitis ↗. Semaglutide is the GLP-1 receptor ↗ agonist in Ozempic and Wegovy; tirzepatide ↗, the Mounjaro and Zepbound molecule, hits the GLP-1 receptor and the GIP receptor at once. Neither is approved for addiction, and this study does not change that.

What it does is add another entry to a growing pile of observational readouts that point the same direction while sharing the same weakness. All of them compare people who chose, or were prescribed, a GLP-1 drug against people who were not. The randomized trials that would settle whether the receptor itself lowers overdose risk have not reported. Until they do, the honest answer is the one the comparator split forces: it depends what you measure against.