Does Semaglutide Reduce Alcohol Craving? Two Trials Disagree

Two randomized trials have now tested semaglutide in people with alcohol use disorder. Drinking fell in both. Craving is where they part company, and the newer of the two made craving its primary endpoint and did not hit it.

That split matters more than a headline can carry. Craving and drinking are different things to measure, they are measured in different ways, and a drug can move one without moving the other.

Whether drinking is safe alongside the drug is a separate question, and our guide to alcohol and peptides covers that one. The question here is whether the drug changes the wanting.

What did the 2026 oral trial set out to prove?

The 2026 oral trial set out to show a drop in cue-elicited craving, measured in a laboratory at a fixed point six weeks in. Its primary outcome is stated plainly: "The primary outcome was laboratory-based alcohol cue-elicited craving at week 6."¹

It ran as a phase 2 double-blind randomized parallel-arm trial in "Fifty individuals with moderate to severe AUD," randomized to oral semaglutide "(3 mg/day for 4 weeks, then 7 mg/day for 4 weeks) or placebo for 8 weeks."¹ Oral and injectable semaglutide are not interchangeable at equal numbers, and our comparison of the GLP-1 pills sets out how the oral form differs. Participants were treatment-seeking, described in the paper as adults with AUD who wanted to reduce or stop drinking.¹²

Baseline drinking was heavy. The publisher's own summary puts it at about six drinks per day and about five heavy drinking days per week, with nearly all participants at high or very high WHO risk drinking levels.²

Cue-elicited craving was a defensible choice for that endpoint, not an odd one. Craving was the measure that had produced a significant exploratory result in the earlier injectable trial. Promoting one trial's promising side finding to the next trial's registered primary is how the field is supposed to work. It is also how a signal gets tested hard enough to fail.

Did the oral trial meet that endpoint?

No, the oral trial did not meet that endpoint. Its abstract is direct about it: "Semaglutide did not significantly reduce laboratory-assessed craving or drinks per day compared with placebo."¹

Two measures failed there, not one. Cue-elicited craving was the registered primary. Drinks per day failed alongside it. A trial built to show that the drug blunts the pull of alcohol in a controlled setting did not show that.

A missed primary endpoint does not mean the drug does nothing. What failed is the specific thing the investigators committed to in advance, in the specific way they committed to measuring it, across 50 people over eight weeks.

A missed primary also changes how much weight the rest of the results can carry. When the registered primary fails, secondary findings are formally no longer protected against the problem of testing many things at once, because the statistical plan that controlled for that was anchored to the primary. A list of significant secondary outcomes appears in the abstract without describing any correction for multiple comparisons, and I could not read the methods section to see whether one was applied.¹

What moved in the oral trial instead?

Heavy drinking days moved instead, with a reported effect of "significantly reduced heavy drinking days (b=-0.580, 95% CI=-1.012, -0.148)."¹ That was a preregistered secondary outcome.

Significant reductions are also reported in drinks per drinking day, naturalistic craving, cannabis use days and alcohol-related consequences, plus more participants lowering their WHO risk drinking level than on placebo.¹ The publisher's summary puts that last figure at 81% of the semaglutide group dropping at least one risk category against about 54% on placebo, and reports no statistical test for the comparison.²

Its lead author frames the result in terms of harm reduction instead of abstinence, saying the study "suggests oral semaglutide may help reduce heavy and harmful drinking" and adding that "even reducing heavy drinking can lead to meaningful improvements for patients and families."² That is a narrower claim than treating alcohol use disorder, and it is the claim the data supports.

Note what happened to craving across those two paragraphs. Measured with cues in a laboratory, it did not move. Measured as everyday craving reported by participants in their own lives, it did. One trial and one group of people produced opposite answers on two instruments.

What did the 2025 injectable trial measure?

The 2025 injectable trial measured alcohol actually consumed in a laboratory session instead of craving. The primary outcomes were grams of alcohol consumed and peak breath alcohol concentration in a self-administration session after treatment.³

That trial was also phase 2, double-blind, randomized and parallel-arm, run at an academic medical center between September 2022 and February 2024. Of 504 people assessed, 48 were randomized, 24 per group, across nine weeks of outpatient treatment with a week-10 discharge visit.³ Participants, investigators and outcome assessors were all blinded, and the design was a hybrid of clinic visits and a laboratory self-administration session.³ Dosing ran at 0.25 mg per week for weeks one to four, 0.5 mg per week for weeks five to eight, and a flexible 1.0 mg at week nine.³

Two details about that design deserve more attention than they usually get. Screening was steep: 504 people assessed to randomize 48, a heavily selected sample and a limit on how far the result generalizes.³ And the primary outcome was measured at the 0.5 mg weekly dose, not at the 1.0 mg offered in week nine, so the headline effect comes from a middle dose rather than the top one.³

One further difference from the oral trial decides a lot of the interpretation. These participants were not seeking treatment, and their AUD was less severe.¹³

Did drinking fall in the injectable trial?

Drinking fell inside the laboratory and not consistently outside it. Grams of alcohol consumed came in at β −0.48 with a 95% CI of −0.85 to −0.11 and P = .01, and peak breath alcohol concentration at β −0.46, 95% CI −0.87 to −0.06, P = .03.³ Each primary outcome separated from placebo.

Outside the laboratory the picture is mixed, and the registered secondary endpoint is the one that failed. Drinks per calendar day came in at β −0.27, 95% CI −0.63 to 0.09, P = .17.³ Drinking days versus abstinent days did not move either, at β 0.90, 95% CI 0.73 to 1.12, P = .89.³

Drinks per drinking day did move, at β −0.41, 95% CI −0.73 to −0.09, P = .04, and heavy drinking days showed a significant treatment-by-time interaction described in the paper as indicating "greater reductions in heavy drinking days over time in the semaglutide group relative to the placebo group."³

Two further exploratory results point the same way. Mean breath alcohol concentration across the laboratory session came in at β −0.48, 95% CI −0.87 to −0.09, P = .02.³ And the proportion of participants with no heavy drinking days rose in the semaglutide group between weeks one to four and weeks five to eight, reaching significance at z −2.93, P = .003, with no other comparison on that measure reaching significance.³

So the pattern inside this one trial is consistent: laboratory drinking down, amount per drinking occasion down, heavy drinking down over time, and the plain count of drinks per day in ordinary life unchanged.

Was craving a real endpoint in the 2025 trial?

Craving was exploratory there, not registered as a secondary outcome. Its registered secondary outcomes are given as drinks per calendar day and cigarettes per day, and treats the weekly drinking and craving outcomes as exploratory.³

The weekly craving result was significant, at β −0.39, 95% CI −0.73 to −0.06, P = .01.³ It is a real number from a real randomized trial and it points the way the headlines said it did.

It is also an exploratory finding, a weaker claim than a registered endpoint, and the distinction is the entire reason the oral trial was built the way it was. Investigators took a promising exploratory signal, made it the primary outcome of the next trial, and it did not replicate in the laboratory.

Why do the two craving results point different ways?

The two craving results point different ways because they are not measuring the same thing. Cue-elicited craving is an acute response to alcohol cues presented under controlled conditions. Naturalistic craving is what a person reports about their own week. One is a provoked laboratory response and the other is lived experience, and nothing requires a drug to change both.

Three readings fit the pattern and the trials do not separate them.

The laboratory cue paradigm may be insensitive to whatever the drug is doing. Neither paper establishes how well a cue response predicts what somebody does with a drink in hand. That prediction is the thing that matters.

Everyday craving reports may be picking up something real that cues miss, such as a reduced appetite for a drink at the point of decision instead of a blunted reaction to the sight of one.

Or self-reported craving in an unblinded-feeling context may be moving partly with expectation. Participants on semaglutide lost weight and had gastrointestinal effects, so functional unblinding is a live concern in any trial of this drug class. That is a criticism of the design, not of the investigators. The injectable trial blinded participants, investigators and outcome assessors alike.³ Blinding on paper does not stop somebody noticing a 5% weight loss.

Were the two groups of people comparable?

The two groups were not comparable, and the difference runs in a direction that matters. The injectable trial enrolled non-treatment-seeking adults with less severe AUD. The oral trial enrolled treatment-seeking adults with moderate to severe AUD who wanted to cut down or stop.¹³

That cuts both ways for interpretation. A treatment-seeking group is more motivated, and that can compress the gap between drug and placebo because the placebo group also improves. A more severe group has more room to fall. No design here compares the two populations, so this is a reason to be cautious about reading one as a replication of the other, instead of an explanation anybody has established.

Severity differed on a documented scale as well. The oral trial enrolled moderate to severe AUD with nearly all participants at high or very high WHO risk drinking levels, against roughly six drinks a day at baseline.¹² The injectable trial's participants were described as having less severe AUD and were not looking for help with it.¹³ Those are different clinical pictures, and a measure like heavy drinking days has more room to move in the first than the second.

Could the weight loss explain the drinking changes?

Weight loss could explain part of the drinking change, and no result here rules it out. Body weight fell 5.05% with a standard deviation of 3.56 in the semaglutide group against 0.18% with a standard deviation of 2.5 on placebo in the injectable trial, with an interaction of β −0.07, 95% CI −0.08 to −0.05, P < .001.³

Other consumption fell too. Cigarettes per day dropped among the smokers there, at β −0.10, 95% CI −0.16 to −0.03, P = .005, in a subgroup of 13 people.³ Cannabis use days dropped in the oral trial, described by the authors as exploratory in a subgroup of 11.¹²

So several consumed substances moved in the same direction at once, in very small subgroups. That is consistent with a shared reward or motivation pathway, the hypothesis driving this whole research program. It is equally consistent with a general reduction in consummatory behavior that happens to include alcohol. A trial that measured only alcohol cannot separate those, and these did not.

What did not change in either trial?

Several measures did not change in either trial, and they constrain the story. In the injectable trial, HbA1c, blood pressure and depression scores showed no significant medication effects.³ That matters for anyone reading the drinking result as a mood effect, because the depression measure did not move.

Drinking days versus abstinent days did not move there either, at P = .89.³ People on semaglutide were not drinking on fewer days. When they drank, they drank less. Each trial points that way, and it is a narrower claim than abstinence.

One procedural null belongs here. Medication did not predict how much participants engaged with the laboratory task, or how long they delayed before drinking, so the self-administration result is not an artifact of the drug changing how people approached the session.³

Safety was unremarkable in the injectable trial, with adverse effects mostly mild and no serious adverse events, no alcohol interactions and no treatment-related discontinuations recorded.³ The oral trial's safety data is in a full text I could not reach.

How big were these trials?

These trials were small and both are phase 2. Forty-eight people randomized in 2025 and fifty in 2026, over nine weeks and eight weeks respectively.¹³

Those are the sizes at which a treatment signal gets looked for, not confirmed. Effect estimates from trials this size carry wide confidence intervals, and some of the intervals above nearly touch zero. The 2026 heavy drinking day interval runs from −1.012 to −0.148, and the 2025 drinks-per-drinking-day interval from −0.73 to −0.09.¹³

Size also explains the subgroups. The nicotine finding rests on 13 smokers and the cannabis finding on 11 users, and findings from subgroups that small are hypotheses with numbers attached.¹³ The authors on both sides treat them as exploratory. That handling is correct, and it is the part that tends to fall away when the findings get quoted elsewhere.

Placebo was the only comparator in each. So nothing here places semaglutide against anything a clinician would otherwise reach for, and the effect sizes above are measured against no treatment. Nothing compares it against another GLP-1 drug either, and our comparison of semaglutide and tirzepatide covers what separates them on the approved indications.

Do the authors' conclusions match their results?

The conclusions mostly match the results, with one claim that reaches further than the data. The 2026 paper concludes that "These findings confirm previous results among non-treatment seekers with less severe AUD and suggest that continued development of semaglutide for AUD is warranted."¹ Its predecessor concludes by "justifying larger clinical trials to evaluate GLP-1RAs for alcohol use disorder."³

That second conclusion is unobjectionable. The first describes a confirmation, and the question is a confirmation of what. Heavy drinking days reached significance in both trials, so there is a genuine replication in that one measure. But the 2026 primary endpoint failed, and the 2025 registered secondary endpoint for real-world drinking failed, so the two trials' clean wins sit in different places.

Applying the same standard in the other direction: a trial that misses its primary endpoint and hits a preregistered secondary in the predicted direction, replicating an exploratory finding from an earlier trial, is a more interesting result than a failed trial. Describing that as confirmation flattens it.

Is this an approved use?

Semaglutide is not approved for alcohol use disorder. The two teams ask for more work instead of reporting a settled treatment, and the 2026 authors ask specifically for "continued development of semaglutide for AUD."¹

Injectable semaglutide reaches patients as Ozempic and Wegovy, and the oral form as Rybelsus. Neither abstract names a brand, so this page does not attach either trial to a particular product, and the approved indications for those three are something I could not source from the label this session. A drug with an approved indication does not need continued development to establish that indication.

I was not able to source the approved indications from the US label this session. The federal drug label service became unreachable from this environment partway through the work, and I would rather say that than quote an indication from memory. Our guide to what the GLP-1 side effect data shows covers the approved products and their documented effects.

Nothing here amounts to a treatment pathway either. Both trials ran as supervised research with blinded outcome assessment, and the 2026 team's own request is for continued development instead of a change in practice.¹

What has not been measured?

Durability, specificity, instrument choice and comparative effect all remain open, and they are the sharpest edge of this evidence.

Durability is unmeasured. Each trial ran eight or nine weeks. Nothing here says what happens at six months, or after stopping.

Whether the effect is specific to alcohol is unmeasured. Nicotine and cannabis both moved in subgroups of 13 and 11 people, and weight fell substantially, and no design here separates a drinking-specific effect from a general one.

The laboratory-versus-life question is unresolved instead of unmeasured. Two trials have now put a laboratory measure at the center and gotten a split answer, while real-world heavy drinking moved both times. An open methodological question remains, namely which instrument to believe, and our guide to why most peptides have no human evidence covers the general version of trusting a proxy measure.

Comparative effectiveness is untouched too. Placebo was the comparator in both trials.

Common questions

Does semaglutide reduce alcohol cravings?

Did the 2026 oral semaglutide trial fail?

What was the primary outcome of the 2026 trial?

Does semaglutide reduce how much people drink?

How many people were in these trials?

Was craving a main endpoint in the 2025 trial?

Did the 2025 trial meet its primary endpoints?

Which drinking measures failed in the 2025 trial?

Could the weight loss be causing the drinking reduction?

Is semaglutide approved for alcohol use disorder?

Does Ozempic or Wegovy reduce alcohol cravings?

What is a heavy drinking day, and why does it matter here?

Why would a GLP-1 drug affect drinking at all?

Did either trial show improved mood?

What dose did the trials use?

Were there safety problems?

How long did the effects last?

Do these trials mean GLP-1 drugs treat addiction?

Sources

1. Schacht JP, Sakai JT, Raymond K, Shelton R. Oral Semaglutide for Alcohol Use Disorder: A Randomized Clinical Trial. The American Journal of Psychiatry 2026; appiajp20260003. DOI 10.1176/appi.ajp.20260003. Abstract read in full via a library reproduction. Full text not read: the publisher's site refused the request from this environment. Quoted for the primary outcome definition, the stated failure to reduce laboratory craving or drinks per day, the heavy drinking day effect size and interval, the randomization and dosing description, the list of significant secondary findings and the conclusion sentence. Safety, adverse events, dropout and the full methods were not available and are described above as unavailable. ↩1 ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18
2. American Psychiatric Association. Newsroom release on the oral semaglutide alcohol use disorder trial, 29 July 2026. Read at source. A publisher-organization press release, not peer-reviewed text, and the weakest of the three sources here. Used only for figures the abstract does not carry: baseline drinking of about six drinks per day and about five heavy drinking days per week, the WHO risk category shift of 81% against about 54%, and the subgroup size for the cannabis finding. The release gives no effect sizes, no confidence intervals and no p-values, and reports no statistical test for the 81% against 54% comparison. Where it and the abstract overlap, the abstract is quoted instead. ↩1 ↩2 ↩3 ↩4 ↩5 ↩6
3. Hendershot CS, Bremmer MP, Paladino MB, Kostantinis G, Gilmore TA, Sullivan NR, Tow AC, Dermody SS, Prince MA, Jordan R, McKee SA, Fletcher PJ, Claus ED, Klein KR. Once-Weekly Semaglutide in Adults With Alcohol Use Disorder: A Randomized Clinical Trial. JAMA Psychiatry 2025;82(4):395-405. DOI 10.1001/jamapsychiatry.2024.4789. Trial registration NCT05520775. Read at source, full text, except the Discussion, Limitations and Conclusions. The page truncated those on two separate reads. The authors' stated limitations are therefore not reflected above, and a third-party copy of the PDF was not used to work around the publisher's wall. Quoted for the design, the dosing schedule, both primary outcomes, every secondary and exploratory outcome including those that did not reach significance, the body weight and cigarette results, the safety summary and the abstract conclusion. The heavy drinking day coefficient is reported as β 0.84 with a 95% CI of 0.71 to 0.99 and P = .04; its sign depends on model coding set out in a supplementary table that was not readable, so this page quotes the paper's own description of the direction instead of interpreting the coefficient. ↩1 ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18 ↩19 ↩20 ↩21 ↩22 ↩23 ↩24 ↩25 ↩26 ↩27 ↩28