IQ Test Scores by Device: iPhone, Android, Windows and Mac

Nobody was asked about their device. Every page view and event on MyIQtested is sent to our analytics endpoint, which reads the browser’s user-agent header once and keeps only the family of the operating system, the family of the browser, the host app when the page is open inside another app, and coarse version numbers. The raw user-agent string is never stored.

The operating system is found by looking for these words in the user agent, in this order, first match wins: “iphone” or “ipod” (iOS), “ipad” (iPadOS), “android”, “cros” (ChromeOS), “windows nt”, “macintosh” or “mac os x” (macOS), “linux” or “x11”. Each test is then put in a group: iPhone (iOS), Android, Windows PC (Windows), Mac or iPad (macOS and iPadOS) and Chromebook (ChromeOS).

The operating system has been kept on every test completion since 17 Sep 2026, so the data runs from that day to 29 Sep 2026, the last full day (UTC) before the export.

The test has 33 multiple-choice puzzles, each with four options and one correct answer: 20 matrix reasoning puzzles (ICAR-based, from the International Cognitive Ability Resource framework), 6 series and logic, 4 verbal and 3 numerical. Every puzzle must be answered and there is no time limit; the median first attempt here took 8.8 minutes.

The score is a fixed conversion of the number of correct answers: 51 + 3 × correct, with a floor of 55 and a ceiling of 150, so scores move in steps of 3. It is calculated in the taker’s browser and is not an empirical norm. That is why this page reports differences and ranks, never a level.

The test is the same on every device: the same puzzles in the same order, laid out for the screen size. The verbal puzzles are re-authored for each language; the other 29 keep the English structure and answer key.

First attempts

185,649

Adults 18+

84,890 (46%)

Countries listed

15, the rest pooled

Window

17 Sep to 29 Sep 2026

Age

Of the 185,649 first attempts, 48% came from people under 18 (including 13% under 13), 46% from adults and 6% from people who did not give an age. The age comes from the one-tap age card shown after the last puzzle.

Country

The largest groups were Indonesia (49%), United States (13%), Japan (10%), India (8%). The 15 countries with 750 or more first attempts have their own column; the rest, 8% of the data, are pooled as Other countries.

Each step keeps the tests that pass it:

• Test completions finished 17 Sep to 29 Sep 2026 (UTC) 324,088
• and the visitor’s first completion 205,656 −118,432
• and a country recorded 205,604 −52
• and 3 to 60 minutes from start to finish 187,421 −18,183
• and one of the five device groups 185,649 −1,772

Every first attempt also passed these checks, so they removed nobody: not a preview, all 33 puzzles delivered and answered, the untimed mode, a score on the scale.

• First attempts only. 37% of completions were a repeat by the same browser (or the completion event firing twice) and are left out, so each person counts once and nobody has seen the puzzles before. A test counts as a first attempt when that browser has no earlier completion since 23 Jun 2026.
• The time from start to finish must be 3 to 60 minutes: 3 minutes is about 5 seconds a puzzle, and a test left open for more than an hour was not taken in one sitting.
• Mac or iPad is one group because the two cannot be told apart. Since iPadOS 13, Safari on an iPad sends the same user agent as a Mac by default; only 1,650 of the 10,223 tests in this group (16%) identified themselves as an iPad.
• Linux (1,768 first attempts) is left out. 75% of them came from Indonesia, India, the Philippines, Pakistan, Bangladesh and Nigeria, as did 81% of Android and 51% of Windows first attempts, so most are likely Android phones with the browser set to desktop mode, not Linux computers. 4 tests from other systems are left out too.
• In-app browsers are kept, in the group of the device they run on: 26% of iPhone tests, 1.5% of Android and 6% of Mac or iPad were taken inside another app, most often the Google app, LINE and other app webviews. Leaving them out moves any group’s adults figure by at most 0.6 points (iPhone: −0.1 with them, +0.5 without).
• Countries with fewer than 750 first attempts are pooled into Other countries. Age not given pools “prefer not to say”, a skipped age card and missing values.
• Cells with fewer than 5 people are pooled into Other countries before publication (97 cells, 230 people), so no cell can identify one person’s score. The aggregate CSV suppresses those cells instead.
• Any group or subgroup with fewer than 30 people is shown greyed as “n too small” and no interval is drawn.

Every figure is relative. Nothing on this page states how high anyone or any group scored on the scale, only how groups differ from each other and where people rank among people of their own age and country.

Every range is a 95% confidence interval, and every difference we point out holds at 95% confidence after correcting for the number of comparisons (Holm method).

Strata and the standard population

A stratum is one age band within one country. Ten age bands and 16 country columns make the strata. With Adjust for age and country on, the standard population is the current filter’s own stratum mix: w_s is the share of everyone in the filter who is in stratum s. With it off, everyone is pooled.

Gap versus all respondents

Coverage is the share of the standard population that the group has people in; strata where the group has nobody are dropped and the rest re-weighted. Coverage is shown in the table when Adjust is on.

Intervals

The factor (1 − p) is there because the reference group, everyone, contains the group. A variance estimated from fewer than 5 people is replaced by the variance of everyone in the filter. With Adjust off there is one stratum and this is the ordinary two-sample formula.

Effect size

Cohen’s d compares a group with everyone else: d = (mean of the group − mean of the rest) / pooled SD, using the same strata weights when Adjust is on. It is in units of the spread of scores, not in points.

Percentile within own age and country

Scores are grouped in 3-point bins and each bin is treated as a single score value. Within a stratum, everyone who shares a bin is tied and shares the range of ranks that bin covers, so the average of their percentiles is the mid-rank percentile. A person’s percentile is their rank among everyone in the data in the same age band and country, whatever device they used. Strata with fewer than 30 people are left out of the percentile figures.

The line in each hero column is the group’s average percentile, and the median in the table comes from the same distribution. With Adjust on, each group’s distribution is the stratum-weighted average of its per-stratum distributions.

• The people are volunteers who chose to take a free online test. They are not a sample of any population, and country and age mixes are those of the site’s visitors.
• The device is the one the test was taken on, not the only one a person owns or uses most. Someone with an iPhone and a Windows laptop counts in whichever group they took the test on.
• The device is not a random choice. Who takes a test on a phone or a computer, and which phone, differs by age, country, income and where they found the test, and adjusting for age band and country removes only the first two.
• The screen differs: the same puzzles are laid out for a phone or a computer, so part of a gap may come from the screen rather than the people.
• The group comes from the browser’s own description of itself. A browser set to desktop mode, or one that disguises itself, lands in the wrong group, and iPads cannot be separated from Macs.
• The score comes from one unsupervised sitting of a 33-puzzle test. It is not a clinical measure, and its scale is a fixed conversion, not a norm.
• 48% of the data is from people under 18, including 13% under 13, who may read the test differently from adults.
• Country is recorded by the site from the visit; it may not be the person’s nationality or where they grew up.
• Differences are shown with intervals, but an interval covers sampling noise only, not the ways the sample was selected.

Dataset version

2026-09-29.1

Generated

2026-09-30

Tests

17 Sep to 29 Sep 2026

Licence

CC BY 4.0

Change log: 2026-09-29.1, first release. If the data or the method changes, the version changes and this line says what moved. Cite the version you used: myiqtested.com/research/iphone-vs-android-iq.