Caffeine shortened cycling completion time (SMD −0.36, 95% CI −0.57 to −0.15) and raised mean power (SMD 0.29, 0.05 to 0.52) against placebo, with no significant change in heart rate or RPE, in a three-level meta-analysis of 20 randomised crossover trials with 226 participants (207 men); doses of 3 mg/kg or less and of 4-6 mg/kg both shortened completion time (−0.42 and −0.34, not compared statistically), but 65% of trials carried some concerns or high risk of bias and Egger's test flagged publication bias for completion time. Read in full
Main caveat: The authors': small trials, especially in subgroups and the meta-regression; caffeine form, task and habitual intake not controlled.
Wu and colleagues pooled randomised crossover trials that compared caffeine with placebo before or during a cycling performance test, and asked how dose, training status and age change the effect.
From 2359 records across five databases they included 20 trials (2011 to 2024) with 226 participants, 207 of them men; doses ran from a total of 32 mg (mouth rinse) or 70 mg to 6 mg/kg, most as capsules about an hour before.
In a three-level model, caffeine shortened completion time (SMD −0.36, 95% CI −0.57 to −0.15, p = 0.0017; 17 trials, every one in the direction of benefit by the authors' account) and raised mean power (SMD 0.29, 0.05 to 0.52; 14 trials), with no heterogeneity; mean heart rate (0.21, −0.06 to 0.48) and RPE (−0.03, −0.30 to 0.24) did not change significantly.
Doses of 3 mg/kg or less (−0.42, 5 trials) and of 4-6 mg/kg (−0.34, 12 trials) both shortened completion time, and well-trained cyclists showed a similar effect to recreational ones; neither split was tested as a difference. Older participants' completion times improved more (β −0.0501 for mean age, p = 0.02).
Egger's test suggested publication bias for completion time, which the authors say may overstate the effect. They conclude that caffeine about an hour before improves cycling performance and that low doses work about as well as higher ones.
Participants
20 randomised crossover trials, 226 participants (207 men, 91.6%); 6 to 20 per trial; study mean ages 20 to 48 years.
134 well-trained cyclists (59.3%), 82 recreational (36.3%), 10 not reported (4.4%), by the authors' definitions (well-trained: at least 3 years of systematic training, 4-5 rides a week, 8-10 h or more a week; recreational: at most 3 rides a week, under 6 h, no long-term competitive training).
No trial studied women only; 3 trials included women (7, 10 and 2; 19 in all, by the sexes listed in Table 1).
Countries: Brazil, the UK, Australia, New Zealand, the USA, China and Spain.
Protocol
Search: five databases from inception to 1 March 2025 (PubMed 441, Embase 515, Scopus 818, Cochrane Library 318, Web of Science 267 records); 989 screened after duplicates and reviews removed, 46 full texts assessed, 20 included (26 excluded for design 4, quality 14, outcome 8).
Inclusion: healthy people, well-trained or recreational cyclists; caffeine before formal exercise as solution, tablet, mouthwash or capsule against an identical placebo; at least one of completion time, mean power, mean heart rate or RPE; randomised controlled or crossover design.Excluded: clinical populations, Wingate-type all-out sprint tests; the Discussion adds that studies in hypoxia, heat, low glycogen, time-to-exhaustion tests and combined supplements were excluded.
Analysis: SMDs (trivial < 0.20, small 0.20-0.49, moderate 0.50-0.79, large ≥ 0.80); three-level model (effect sizes within studies), REML, cluster variance estimation with the CR2 small-sample correction; Cochran's Q and I²; Egger's test and funnel plots; leave-one-out; pre-specified subgroups by dose (≤3 vs 4-6 mg/kg) and training level; meta-regression on age.
Funding and interests: no financial support; no conflicts declared.
Key Results
| Outcome (caffeine vs placebo) | Trials | SMD (95% CI) | p | I² |
|---|---|---|---|---|
| Completion time | 17 (20 effect sizes) | −0.36 (−0.57 to −0.15) | 0.0017 | 0% |
| Mean power output | 14 | 0.29 (0.05 to 0.52) | 0.02 | 0% |
| Mean heart rate | 11 | 0.21 (−0.06 to 0.48) | 0.11 (not significant) | 0% |
| RPE | 10 | −0.03 (−0.30 to 0.24) | 0.82 (not significant) | 0% |
Subgroups (Table 2; none tested against the other):
| Outcome | ≤3 mg/kg | 4-6 mg/kg | Well-trained | Recreational |
|---|---|---|---|---|
| Completion time | −0.42 (−0.81 to −0.02), 5 trials, n = 57 | −0.34 (−0.59 to −0.10), 12 trials, n = 143 | −0.38 (−0.63 to −0.13), 12 trials | −0.32, printed CI −0.71 to −0.06 with p = 0.09, 5 trials |
| Mean power | 0.27 (−0.20 to 0.73), not significant, 4 trials | 0.29 (0.02 to 0.57), 10 trials | 0.25 (−0.01 to 0.51), p = 0.06, 11 trials | 0.44, printed CI 0.10 to 0.98 with p < 0.10, 3 trials |
| Mean heart rate | 0.41 (−0.01 to 0.82), p = 0.05 | 0.07 (−0.29 to 0.42) | 0.21 (−0.09 to 0.51) | 0.23 (−0.42 to 0.87) |
| RPE | −0.06 (−0.47 to 0.35) | −0.004 (−0.37 to 0.36) | 0.01 (−0.30 to 0.33) | −0.15 (−0.67 to 0.37) |
Age: completion-time benefit grew with mean age (β −0.0501, 95% CI −0.09 to −0.01, p = 0.02); no age effect on power (β 0.0145, p = 0.55), heart rate (0.0154, p = 0.61) or RPE (−0.0021, p = 0.94). Study mean ages ran 20-48, so this is a between-study association.
Sensitivity: leaving out any one trial kept completion time significant (−0.33 to −0.45, p < 0.01); for mean power the estimates ran 0.22 to 0.30 with p = 0.05 to 0.11, so the power result is fragile, although the authors describe all primary outcomes as unchanged in significance.
Publication bias: Egger's test significant for completion time (p < 0.05); not for power (p = 0.12), heart rate (p = 0.41) or RPE (p = 0.80).
Risk of bias (RoB 2): 35% of trials low, 40% some concerns, 25% high; concerns mainly in randomisation, outcome measurement and selective reporting.
Practical Application
Caffeine makes a cycling time trial faster by a small margin: across 20 crossover trials completion time improved by SMD −0.36 and mean power by 0.29 against placebo, and 3 mg/kg or less shortened completion time about as much as 4-6 mg/kg. Trial it in training first, as SYNTH-67 requires; individual response varies and one genotype trial in the base found some men slower on 4 mg/kg (REF-910).
Inference, not tested here: start at the low end. The ≤3 mg/kg subgroup (5 trials) gave a completion-time effect no smaller than 4-6 mg/kg; the two were not tested against each other and the low-dose power result was not significant. SYNTH-67 and SYNTH-62 give 3-6 mg/kg, "possibly effective from 2 mg/kg" in SYNTH-67; this review supports the low end of that range rather than changing it.
Do not expect it to feel easier. RPE and heart rate did not change significantly while output rose; inference, not tested here: on caffeine a rider tends to ride harder at the same perceived effort, so pacing by RPE alone may let the target creep up.
Older riders may gain more (a between-study association with mean age, 20-48); this is not a reason to raise the dose.
The size is likely an upper bound. Egger's test flagged missing null trials for completion time, and 65% of trials had some concerns or high risk of bias.
Limitations
The authors': small trials, especially in subgroups and the meta-regression; caffeine form, task and habitual intake not controlled; mostly short, high-intensity cycling tasks; 91.6% men and no women-only trial; exclusion of special conditions limits generalisation; potential publication bias that may overstate the effect.
Inclusion against what was included (our reading of Table 1): the Discussion says time-to-exhaustion tests and heat were excluded, yet one included trial rode to exhaustion at 80% VO2max and another to exhaustion at 35 °C; Wingate-type sprints were excluded, yet one trial measured five 6-s sprints after 90 min; one trial measured only heart rate during a fixed 90-min ride.
Reporting: completion time is described as 17 studies with Q on 19 degrees of freedom (20 effect sizes); two recreational subgroups print a confidence interval that excludes zero with a p value above 0.05 (completion time −0.32, −0.71 to −0.06, p = 0.09; power 0.44, 0.10 to 0.98, p < 0.10); the dose and training-level conclusions rest on subgroups never tested against each other.
Dose groups: the ≤3 mg/kg group mixes capsules, a solution and caffeinated gum taken after 10 km; two trials gave absolute doses (32 mg mouth rinse, 70 mg tablets) whose subgroup is not stated.
Not read: per-trial risk-of-bias ratings and GRADE certainty, which sit in supplementary files and a figure.
How this entry was checked and changed
Why this level
a meta-analysis of randomised trials whose risk of bias we read in the full text (35% low, 40% some concerns, 25% high), held at A- by small trials (6 to 20 participants each), a positive Egger's test for completion time, and moderator claims built on subgroups not tested against each other