Thanks to Elizabeth van Nostrand and Robert Mushkatblat for comments on a draft of this post.

[Epistemic status: I like investigating things but I have no particular medical expertise. My subjective confidence in this post being correct is around 50%. I started writing this post when I thought the risk reduction was higher; I no longer think it's a slam dunk that you should try this. Insert magic legal incantation about not medical advice blah blah blah.]

TLDR: see title. 4 studies all look roughly consistent with something like 15% risk reduction for recently-vaccinated populations and 40% risk reduction for unvaccinated populations. Metformin is a medication very widely prescribed for diabetes with minimal side effects, so the costs are low. There are some reasons to think results could be better than this, and some reasons to think they could be worse.

Lit review

Here are all the studies I can find that investigate the effects of metformin on rates of long covid, in descending order of quality:

• The ACTIV-6 RCT randomized 2983 people with active COVID symptoms (who'd had them for less than a week) and a positive test result from September 2023 to May 2024: 1544 on placebo and 1439 on metformin (500mg day 1, 1000mg days 2-5, 1500mg days 6-14).
  • At 90, 120, and 180 days after, they measure (1) if patients self-report nonzero COVID symptoms (and didn't reacquire covid within the past month) and (2) if a physician has diagnosed them with long covid.
  • Self-reported long COVID symptoms 180 days out went down by 21% (0.037 to 0.029)^[1], and clinician diagnosis of long COVID went down by 51% (0.014 to 0.007). They report this as a negative result because it didn't meet their preregistered threshold for efficacy; CIs slightly overlap zero.
    • Note that while the 50% number is their preregistered secondary outcome, it's the best-looking number out of everything they measure and probably got a bit lucky; see here for a plot of a few different confidence intervals. Everything looks good for metformin (and generally looks better the further out you look) but 50% relative risk is an outlier. And again, the confidence intervals are wide here.
  • Frustratingly they don't do any subgroup analyses and haven't shared raw data. I've emailed the primary contact for the study asking how effects vary with vaccination status but haven't heard back yet.
  • IMO this study is pretty good. They preregistered everything, they have decent sample sizes, they aren't doing the publication bias thing because they report this as a negative result, etc. I have no complaints except that I want subgroup analyses.
    • The same people running this thing tried some other substances in RCTs too: they found no effect for ivermectin, fluvoxamine, fluticasone, or montelukast. (I think actual no effect rather than "fairly good effects with wide CIs" in all of these.) So they're not generally predisposed to find positive results. 
       
• COVID-OUT was an RCT from Dec 2020 to Jan 2022 on 1323 people, of whom  1126 answered a survey 6 months out. They filtered for adults age 30-85 who were overweight (median 45yo, 29.8 BMI), so a higher-than-baseline risk population, and required symptom onset within the past 7 days. Split 50/50 into placebo vs metformin; exact same dosing schedule as the previous trial (in fact this one came first and I think ACTIV-6 is copying it). Average first dose was 5 days after symptom onset.
  • They find a 41% reduction in the risk of long COVID (as measured by patient self-report of a diagnosis from a medical provider), from 10.4% in the placebo group to 6.3% in the metformin group.^[2] In the subset who started metformin within 3 days of symptom onset, they saw a 63% reduction.
  • In the subset of vaccinated patients (about half the study population), they find a 15% decrease from 7.2% to 6.1%, but with fairly wide confidence intervals overlapping zero (they caution that it wasn't powered to detect subgroup effects though). Among the 57 participants with a booster vaccine, only one reported long COVID at all.
    • Note that this means the unvaccinated effects are even stronger: something like 14.3% to 6.5% risk from placebo to metformin, for a 54% reduction. (And it suggests that vaccination is itself a 2x risk reduction.)
  • A second paper about the same trial looks at short-term effects of metformin on viral load (I think with only n=775 of the 1323 participating through day 10) and hospitalization.
    • They find a 3.6-fold viral load reduction (only p=0.027 which is kinda surprising to me with effect sizes that big?) and 28% less likely to have a detectable viral load at day 5 or day 10.
      • This is a kinda crazy large effect, and seems sort of implausible given fairly mild effects elsewhere? Not sure how to reconcile this well; maybe the highish p-value here is somehow related, and this measurement was kinda noisy?
    • Hospitalization or death at 28 days is 58% lower in the metformin group.
    • They see like 2x larger effects on the unvaccinated subpopulation, and say "Effective primed memory B- and T-cell anamnestic immunity prompting effective response by day 5 in vaccinated persons may account for this trend in both trials".
  • Reassuringly for the quality of the study, they also look at ivermectin and fluvoxamine and don't find positive effects.
• Johnson et al (2024) looks at some correlational data in the National COVID Cohort Collaborative (N3C) and finds that among patients with Type 2 diabetes, the ones who were taking metformin had 21% lower rates of long covid diagnosis or death (2.0% to 1.6%) than patients taking other diabetes medications, and 15% lower rates of post-covid symptoms as measured by "computable phenotype" which I think means they look at your logged symptom data and guess if you have long covid. p<0.001 for both but note this is correlational so add salt to taste.
  • They also look at data from PCORnet, which has wider confidence intervals and worse results (13% decrease by diagnosis, 4% increase by computable phenotype). They think that fully a quarter of the PCORnet population has long covid as measured by computable phenotype though, so I'm more skeptical of this patient population and whatever they're doing to infer things.
  • Here's a random slide that shows the confidence intervals in a little plot.
  • If you download the supplemental figure PDF from the first link in the parent bullet, Figure 5 breaks things down by time after covid infection, which is nice, but it looks like fully 5% of their study populations in both cohorts died within 6 months?? Type 2 diabetes is not a very fatal condition and neither is covid! I have no idea what's going on there and assume I'm misreading the plot or something. But this makes me trust the study less.
• The TOGETHER trial in Brazil in 2021 randomized patients to 10 days of 750mg extended release metformin twice daily, starting within 7 days of symptom onset, in populations with at least one risk factor or over 50 years of age. Sample size n=418.
  • They look at effects on hospitalization for >6h and find basically no effect (sign of the effect varies depending on exactly how you filter), though the credible intervals on relative risk are super wide, like 0.6 to 1.7. They do not investigate long COVID.
  • This is in pretty direct conflict with the results of the COVID-OUT trial which does quite similar stuff on hospitalization; the COVID-OUT authors note this discrepancy and dedicate a paragraph in their discussion section to it, which I've attached in a footnote.^[3] I find their response kinda weak.
  • This study has since been retracted, I believe for some errors in their early stopping criterion analysis. From a cursory investigation I haven't been able to figure out exactly what the deal is or what a corrected estimate of the final endpoint would be; it sounds like the data wasn't fraudulent or anything, they just had some analysis mistakes. But given this I'm not putting much of any weight on their conclusions.

Does this only matter for unvaccinated people? 

In my opinion the above results present a pretty compelling case for use of metformin in unvaccinated subpopulations, but I expect that very few readers of this post in 2025 fall into that category.^[4]

We have evidence from COVID-OUT that the effects are lessened in vaccinated subpopulations, which makes some intuitive sense. The other studies don't weigh in on this question.

Some unfounded speculation on this topic:

• Conditional on real effects for unvaccinated populations (which look fairly robust), it would be kind of surprising if vaccination brought the effects to zero, so probably it does something.
• If we take for granted the numbers in COVID-OUT for lack of anything better to go on, we probably expect a 2-3x reduction in efficacy for vaccinated subgroups.
• COVID-OUT was Dec 2020 to Jan 2022, so probably most vaccinated people in that trial were pretty recently vaccinated. It seems reasonable to guess that recent vaccinations do more than old vaccinations, so if you have not been vaccinated in over a year, my wild guess would be that you're at more like a 1.5x reduction in efficacy related to unvaccinated numbers.
  • On this model, one would expect that the later-running ACTIV-6 trial would see smaller differences between vaccinated and unvaccinated populations, because their vaccinated group would have gotten their shots longer ago on average. If I hear back from the authors of that study we can award or subtract Bayes points accordingly.

Personally, as a person who hasn't gotten a vaccine in over a year, I plan to take metformin if I get covid, but if I didn't already have it on hand, I would probably prioritize "get a Novavax shot with few side effects" over "acquire metformin" as a low-cost intervention. (And I should really get around to getting a booster shot sometime.) But you could do all of these things!

Mechanism of action

The COVID-OUT trial found acutely lower viral load, which they say "may explain the clinical benefits in this trial", but they also say "Metformin is pleiotropic with other actions that are relevant to COVID-19 pathophysiology". They also say "An improved effect size for clinical outcomes when therapies are started earlier in the course of infection is consistent with an antiviral action", and talk a bit about why they thought to try it in the first place:

The selection of metformin was motivated by in silico modeling, in vitro data, and human lung tissue data that showed that metformin decreased SARS-CoV-2 viral growth and improved cell viability [2–4]. The in silico modeling identified protein translation as a key process in SARS-CoV-2 replication, similar to protein mapping of SARS-CoV-2 [3]. Metformin inhibits the mechanistic target of rapamycin (mTOR) [5], which controls protein translation [6, 7]. Metformin has shown in vitro antiviral actions against the Zika virus and against hepatitis C via mTOR inhibition [8–11].

Wikipedia says that metformin has anti-inflammatory effects, which the COVID-OUT paper also notes briefly as a possible mechanism. Overall it seems not very settled.

How do you get metformin? How risky is it? What dosing schedule should you use?

It's a prescription drug in the US, but an extremely common one. You can try your luck talking to a physician upon developing symptoms, but since it seems like starting early is particularly valuable I would prefer to have some ability to start ASAP and acquire it ahead of time. 

Metformin seems extremely low-risk; the RCTs linked above actually say things like "metformin is safe" without caveats. They even let pregnant women into the COVID-OUT study! Severe kidney disease is the only major contraindication I see from some searching. Some websites warn about lactic acidosis but it seems like this is a non-issue.

The two big RCTs that find good effects use immediate release metformin with 500mg day 1, 500mg twice a day for days 2-5, and 1500mg from days 6-15, but it's not like people tested a lot of options as far as I can tell. The authors of the COVID-OUT have some stuff to say on dosing in the Discussion section:

The magnitude of metformin's antiviral effect was larger at day 10 than at day 5 overall and across subgroups, which correlates with the dose titration from 1000 mg on days 2–5 to 1500 mg on days 6–14. The dose titration to 1500 mg over 6 days used in the COVID-OUT trial was faster than typical use. When used chronically, that is, for diabetes, prediabetes, or weight loss, metformin is slowly titrated to 2000 mg daily over 4–8 weeks. While metformin's effect on diabetes control is not consistently dose-dependent, metformin's gastrointestinal side effects are known to be dose-dependent [25]. Thus, despite what appears to be dose-dependent antiviral effects, a faster dose titration should likely only be considered in individuals with no gastrointestinal side effects from metformin.

So it kinda sounds like you could yolo a faster titration if you're up for some nausea/diarrhea risk? I am not a doctor!

COVID-OUT found substantially stronger effects by starting metformin within 3 days of symptom onset instead of 5. If you're quick with a covid test or proactively take metformin when you start having covid-like symptoms, you might be able to get large wins here compared to the studies above.

Reasons to be skeptical

• On priors most weird medical interventions that haven't become standard practice are bunk, especially if they have wide confidence intervals. I think this should be around a 2:1 update against this working; that still leaves me feeling like it's probably worth it, but I think it's reasonable to decide it isn't worth it by your lights if you have slightly different epistemics or values.
• The RCT with the strongest results was run on an overweight and older population; maybe all the effect here is from mitigating really serious infections for high-risk patients (which in turn reduces long covid) but it does ~nothing for young healthy populations.
• Carolyn T. Bramante is the lead author on both the ACTIV-6 and the COVID-OUT trial, so the author populations aren't independent.
  • She's a professor of medicine at the University of Minnesota though, it's not like these studies are run by fringe cranks.
• Taking drugs is generically bad for Chesterton's fence reasons.
  • I agree; the badness of long covid outweighs this in my utility function but ymmv.

1. ^{^}

   These aren't the rates reported in the main body of the study, but the rates they report use a denominator of all study participants instead of the subset that responded to surveys 180 days out, and the latter seems like a better metric especially for getting accurate raw rates. It doesn't affect the relative numbers much though. Concretely 180-day self reports went from 46/1251 to 33/1139 and 180-day clinician diagnosis went from 18/1251 to 8/1138 (idk what happened to the one metformin participant who only answered the self-report question).

2. ^{^}

   I'm a bit confused about these numbers actually, because they say the sample sizes are 562 placebo and 564 metformin, but you can't get those percentages by rounding any possible patient numbers here; 58/562=10.320% but 59/562=10.498%, and 35/564=6.205% but 36/564=6.383%. The only way I can get the ratio of these numbers to round to their stated hazard ratio of 0.59 is by taking 59 placebo and 35 metformin. Perhaps they're silently doing something like what I did in the previous footnote? Paper authors should show lots of significant digits! 

3. ^{^}

   "Conversely, an abandoned randomized trial testing extended-release metformin 1500 mg/d without a dose titration did not report improved SARS-CoV-2 viral clearance at day 7 [20]. Several differences between the Together Trial and the COVID-OUT trial are important for understanding the data. First, the Together Trial allowed individuals already taking metformin to enroll and be randomized to placebo or more metformin [20, 21]. To compare starting metformin versus placebo, the authors excluded those already taking metformin at baseline and reported that emergency department visit or hospitalization occurred in 9.2% (17 of 185) randomized to metformin compared with 14.8% (27 of 183) randomized to placebo (relative risk, 0.63; 95% confidence interval, .35 to 1.10, Probability of superiority = 0.949) [22]. Thus, the Together Trial results for starting metformin versus placebo are similar. Second, 1500 mg/day without escalating the dose over 6 days would cause side effects, especially if the study participant was already taking metformin [23]. Third, extended-release and immediate-release metformin have different pharmacokinetic properties. Immediate-release metformin has higher systemic exposure than extended-release metformin, which may improve antiviral actions, but this is not known [24, 25]. Given the similar clinical outcomes between immediate and extended-release, a direct comparison of the 2 may be important for understanding pharmacokinetics against SARS-CoV-2."

   I don't know where they're getting some of these numbers like the 9.2% vs 14.8%, I don't see them in the text of the TOGETHER study? Or the stuff about their policies for patients already taking metformin.

4. ^{^}

   Might still be useful to keep in mind for unvaccinated relatives though!