The most legible font, measured.

This started as a bad afternoon. We had just shipped a beautiful headline serif across our whole app, the kind of font you pick because it makes screenshots look like a magazine, and then one of us glanced at a lock-screen widget from across a desk and could not read it. Not “squint a little” could not read it. Could not read it.

The normal move here is to shrug and pick Helvetica. Instead we got curious about a question that turned out to be surprisingly hard to answer with a straight face: which font is actually the most readable on a screen? Everyone has an opinion. Verdana people, Inter people, the dyslexia-font crowd, your one friend who says serifs are a psyop. What almost nobody has is a number, because running human reading studies across dozens of fonts costs a fortune, and the studies that exist test three or four faces at a time.

So we built the next best thing we could actually afford: a machine that reads the way screens fail. It renders the same words in every font, then blurs them, shrinks them, tilts them, inverts them, buries them in visual noise, packs them too tight, and washes out their contrast, and after each insult it asks an OCR engine to read what is left. Fonts that keep their letterforms keep their scores. Fonts that dissolve, dissolve measurably. Our beautiful headline serif came in dead last, twenty-fourth of twenty-four, and the three fonts we ship today exist because of everything else this machine taught us.

Reading has mechanical preconditions before it has psychology. A reader, human or machine, must identify a letter at the angular size it lands on the eye, separate it from its neighbors, discriminate it from its lookalikes, and do all three through real optics: blur, contrast loss, inversion, projection, clutter. The battery measures each precondition directly. The size sweep is identification. A new tight-set condition is separation, because crowding, not acuity, is usually what limits reading. Adversarial codes plus an empirical confusion matrix are discrimination. The degradation cells are the optics.

And no, nobody reads 8-pixel text all day.Nominal size is the physical model of distance. Sixteen-pixel text read from twice as far subtends the same visual angle as eight-pixel text from here: the phone across the desk, the lock screen at arm’s length, the sign down the hall. The rasterized 8 px version also carries fewer pixels than distant 16 px text, so the test is strictly harder than distance viewing, which makes every threshold on this page conservative. A font’s small-size threshold is its distance budget. What this index still does not measure: reading speed, comprehension, comfort, taste, or anything about any human population. Machine legibility, stated exactly.

The setup is simple enough to explain over coffee. We render the exact same words in every typeface, at the same sizes, then damage every image the same way, and ask Apple’s Vision text recognizer, the thing behind Live Text on your iPhone, to read it all back with its dictionary turned off. The score is just how many characters it got right. Eight runs per cell so no font gets lucky, identical images and shuffle seeds for everyone, and every raw run published so you can catch us if we slip.

The twelve conditions: body text at 9 and 12 px, both under blur, a hard downscale, a keystone warp, inverted polarity, a bright cluttered background, adversarial codes straight and downscaled, and the two v3.1 additions: crowded, body text at tight leading and tracking, and low-contrast, gray-on-white the way interfaces actually style secondary text.

The second number:a size sweep from 6 to 16 px gives each font its “legible down to” figure, the smallest size holding 90% accuracy. Ranks blur inside error bars; thresholds do not.

Read it in tiers.T1 marks every face whose confidence interval overlaps the leader’s, the rule set before the table was made, and by that rule Verdana stays in the tier: its wide interval is itself a finding, driven almost entirely by the crowding cell. The profiles separate what ranks cannot: Legibility Sans holds the deepest small-size floor, Plex and Atkinson Next take the crowding cell, Kept Sans carries a brand while keeping tier, and Verdana wins clean 9 px text outright while reading nine points under the tier at tight setting.

If you came here from a search box, here is the straight answer, with the caveats attached where you can see them. By our measurements, the most legible screen fonts are a six-way top tier: IBM Plex Sans and Public Sans at 96.5%, our own Legibility Sans at 96.4%, Kept Sans at 96.3%, Atkinson Hyperlegible Next at 96.0%, and Verdanaat 95.6%. The differences inside that tier are smaller than our error bars, so anyone who tells you one of them is “the winner” is selling something. What we can say precisely: if your text gets tiny, Legibility Sans held 90% accuracy at 8 pixels and nothing else did; if your text gets dense, avoid Verdana, which loses nine to twelve points to every other tier face when lines are packed tight.

And a distinction worth thirty seconds of your life: researchers use readability for how hard a text is to understand and legibility for whether you can make out the letters at all. We measure legibility, with a machine, and we say so everywhere, because the fastest way to lose your trust would be to pretend a robot can tell you what is comfortable for your eyes. What it can tell you is which letters survive being small, blurry, crowded, and washed out, and that turns out to be the part a font can control.

Every famous legibility project we could license sits in this table, measured under the same rules. The results deserve plain statement, with respect: these are serious projects, and several of them taught us the moves we used.

Atkinson Hyperlegible Next, 96.0, top tier.The Braille Institute’s engineering is the real thing; it is the chassis Legibility Sans is built on, and its crowding cell (99.3) is the best in the index. IBM Plex Sans and Public Sans, 96.5.The quiet finding of this run: a corporate house font and the U.S. government’s default are world-class screen faces, and nobody markets them that way. Lexend, 94.7. Built around reading-proficiency research; strong mid-table here, hit hard by tight setting (84.8). Its claims are about human reading speed, which this index does not measure, so the two results do not contradict.

Comic Sans MS, 93.7. The folk recommendation for dyslexic readers measures respectably at clean sizes and collapses when crowded (66.5). OpenDyslexic, 92.3.The machine result, stated flat: solid at clean sizes, 58.6% under 9 px blur in the v3 archive. Separately, controlled human studies (Wery & Diliberto 2017; Kuster et al. 2018) find no reading advantage for it; our machine number neither confirms nor contradicts them, because the two measure different things. Andika, 93.3. SIL’s literacy face is drawn for beginner readers at generous sizes, and its crowding cell (60.8) says exactly that: use it the way it was designed to be used, spacious and large. None of these results is a claim about any human reader; they are what a machine recovers under stress, published so the projects themselves can rerun and argue.

The holdout.Our fonts were tuned against this benchmark, so twenty words and four codes were held out of every design loop and run once against final binaries. The test of overfitting is the gap between a font’s main-corpus and holdout cells: for every face of ours that gap is small and runs in both directions, which is what honest generalization looks like. The confusion matrix.New in v3.1: the reader’s actual substitution errors, accumulated per font from the alignment. The residual errors are the same three pairs for every strong face, I to 1, O to 0, l to 1, which is what the floor of this task looks like, and they match the human confusion literature. Pseudowords and matched x-height carry over from v3 in the dataset: scrambling the lexicon costs 0 to 9 points, typically 2 to 4, and no font changes tier; equalizing x-height converges the sturdy faces and leaves the fragile ones fragile.

Nobody has earned “the most legible font, period,” including us. Published research finds different fonts win for different human readers, so a universal human ranking is not even coherent. What we have is narrower and more useful: a reproducible machine index anyone can run, extend, or attack, a protocol precise enough to hand to your own AI and rebuild from scratch, and a public repo with the peer-review-grade roadmap already written down.

If you run it and beat us, open an issue. We will update the table and say so.

Kept Sans, Numen Title, and Legibility Sans are free and open under the SIL Open Font License 1.1. The zip has every weight in TTF and WOFF2; the repo has the harness, the OCR tool’s source, the corpus, and every raw number on this page.