Safari 27.0 shipped on September 17, 2026. The interesting number is not 80 features, it's the gap: WebKit's own post says "The number of features alone rose from 58 to 80 since the first beta in June." The resolved-issue count went from the 525 announced at WWDC to 844.
I wrote about the beta's 58 features on September 14 and named six that would change how you build an interface. Three days later, on September 17, the release was out, which makes this the cheap moment to check my own homework: did the six survive, and what arrived in the 22 that came after?
Quick take: the six from June all shipped, unchanged. The genuinely new work for interface people is scroll anchoring through
overflow-anchor, four additions to the colour functions includingalpha(), an unclampedprogress(), and:host:has(). The number worth quoting to anyone who still calls Safari a laggard is 844.
What Arrived After the Beta?
Twenty-two features, and they cluster in a way the release post's alphabetical sections hide.
| Area | New since June |
|---|---|
| CSS colour | alpha(), multi-colour color-mix(), image(<color>), light-dark() taking images |
| Scrolling | overflow-anchor, defaulting to auto |
| CSS values | progress() gains no-clamp |
| Selectors | :host:has(), case-sensitive s on attribute selectors |
| Containment | contain: style for CSS quotes |
| Canvas and SVG | optional radii on roundRect(), <use> referencing external files without a fragment |
| Tooling | Safari MCP, a server for coding agents |
Four of those seven rows are things you type into a stylesheet. Compare that with the beta, where two rows out of six were stylesheet work. For a point release between beta and stable, that's an unusual amount of authored CSS.
Why Will You Notice Scroll Anchoring Without Reading About It?
overflow-anchor is the whole scroll-anchoring feature, and it ships defaulting to auto. In plain terms: when something above your reading position finishes loading and takes up space, the page no longer shoves the paragraph you were reading down past your eyes.
Every one of us has built the bug this fixes. An image without dimensions, an ad slot that resolves late, a font swap that reflows a heading into two lines, and the reader loses their place. The usual defence was to reserve space with aspect-ratio and be disciplined about it, which works right up until a third-party embed ignores you.
Where it earns its keep is the phone, because that's where a 60px shift is half a paragraph rather than a line. WebKit's post describes the effect the same way: scroll anchoring keeps content from jumping when something loads above the viewport. You can check it on any page of your own by putting an image without width and height every few paragraphs, scrolling past the first one and letting it load.
/* Only when the browser's anchoring fights your own scroll maths. */
.virtual-list {
overflow-y: auto;
overflow-anchor: none;
}
That's the only declaration most sites will ever need, and the reason to keep it narrow. overflow-anchor: none on a virtualised list is correct, because you are managing scrollTop yourself and two managers is one too many. The same declaration on body is a regression you shipped on purpose.
Which Four Colour Additions Matter, and Which Rewrites Tokens?
alpha() is the small one that will show up in every design system inside a month:
:root { --brand: oklch(55% 0.21 258); }
.overlay {
/* Before: restate every channel you are not changing. */
background: oklch(from var(--brand) l c h / 80%);
/* After. */
background: alpha(from var(--brand) / 80%);
}
The formal grammar is alpha( [from <color>] [/ [<alpha-value> | none]]? ), and the reason it matters isn't brevity. The full relative syntax makes you name a colour space and restate the channels you're leaving alone, and each of those is a chance to shift a hue by accident. alpha() keeps the origin colour in its own space and changes exactly one channel. A token file that needs a 12%, a 40% and an 80% variant of nine brand colours goes from a wall of conversions to twenty-seven honest lines.
color-mix() now takes more than two colours, which is why the middle swatch in the preview above has three. That lands differently than it reads: a two-colour mix is a blend, a three-colour mix with explicit percentages is a small palette definition, and you can now write a tertiary directly rather than nesting mixes inside each other.
The other two are narrower. image(<color>) lets a solid colour stand where an image is expected, and light-dark() now accepts image values, so a scheme-aware background image stops needing a media query:
.panel {
background-image: light-dark(url("/paper-light.avif"), url("/paper-dark.avif"));
}
If you're composing tokens rather than reading about them, the palette generator below produces the OKLCH ramp these functions expect, and the swatches it gives you paste straight into an alpha() origin.
What Escape Hatch Does progress() Grow?
progress(value, start, end) returns how far a value sits between two bounds, clamped to 0 through 1. no-clamp removes the clamp:
/* Clamps at 1, so the effect flattens past the boundary. */
--lift: progress(300, 0, 100);
/* Returns 3, so the effect keeps going. */
--lift: progress(no-clamp 300, 0, 100);
The keyword goes first and takes no comma after it. Why care? Because clamping is what makes scroll-driven motion feel like it hits a wall. An overscroll bounce, a parallax layer that should keep drifting past a section, a shadow that should keep growing while you pull: all of those want extrapolation, and all of them were previously faked with a second calc() that undid the clamp.
My rule of thumb: unclamp the source and clamp the consumer. An unbounded custom property will eventually meet an input nobody sanity-checked, and a translate of 4000px is harder to debug than a flat animation.
How Does :host:has() Close a Component Hole?
:host:has() lets a web component style itself from the state of its own children:
:host:has(input:invalid) {
border-color: oklch(58% 0.18 22);
}
Before this, a component that wanted to react to its own content either exported a part and asked the page to style it, or set an attribute from JavaScript on every state change. Neither is wrong. Both mean the component's appearance lives somewhere other than the component.
The attribute-selector change belongs beside it. Attribute matching can now be made explicitly case-sensitive with an s modifier, the mirror of the i modifier we've had for years, which matters for anything keyed on a value where case carries meaning.
Which Three Smaller Changes Will Somebody Hit?
contain: style now applies to CSS quotes. That reads like nothing until you have a component that sets quotes and gets nested inside another that also sets quotes: the counter for nesting depth is document-wide, so the inner component picks up the outer one's level and shows the wrong glyph. Style containment scopes that counter to the container. It's a bug fix for a problem most people worked around by not using quotes at all.
roundRect() on a canvas context now takes its radii argument optionally, so ctx.roundRect(x, y, w, h) draws a plain rectangle instead of throwing. Small, but it removes an if from every wrapper anybody wrote around it.
And SVG <use> can reference an external file without a fragment identifier. <use href="icon.svg" /> works where you previously had to write icon.svg#root and therefore had to put an id on the root of every icon you shipped. If you generate a sprite sheet, nothing changes. If you ship one file per icon, you just deleted a build step.
What Would I Adopt Today, and What Would I Wait On?
Shipping order matters more than the feature list, and the cross-browser picture is better than the "Safari catching up" story suggests. web.dev's September 2026 platform roundup lists alpha() in Chrome 151, Firefox 155 and Safari 27, which it calls Baseline Newly available, and progress() with no-clamp in Firefox 155 and Safari 27 (it does not say Chrome has it). overflow-anchor is new only in Safari: Chrome 56, Edge 79 and Firefox 66 had it already. I could not confirm cross-browser status for multi-colour color-mix() or :host:has(), so treat those as Safari-only until you have checked. My split:
| Feature | Other engines (per web.dev, Sept 2026) | Ship now? | How |
|---|---|---|---|
overflow-anchor | Chrome 56, Edge 79, Firefox 66 already | Yes, by doing nothing | It's a default; the only work is none where you manage scroll yourself |
alpha() | Chrome 151, Firefox 155 | Yes, with a fallback | Declare the opaque colour first, then the alpha() line |
color-mix() with 3+ colours | Not confirmed | Yes, same pattern | A browser that parses only two colours drops the declaration |
light-dark() with images | Not confirmed | Yes | The media-query version is a two-line fallback, not a rewrite |
progress() with no-clamp | Firefox 155 (Chrome not stated) | Careful | The unclamped value is unbounded, so clamp where you consume it |
:host:has() | Not confirmed | Yes inside a component | Shadow DOM already forces you to fall back gracefully |
| Safari MCP | Safari only | No | Single vendor, unproven protocol |
The pattern for every "yes with a fallback" row is the same two lines, and it's the pattern the preview at the top of this page is built on:
.chip {
background: #2563eb;
background: alpha(from #2563eb / 80%);
}
An engine that does not know alpha() parses line one, fails to parse line two, and keeps line one. That's the cascade doing exactly what it was designed for. The fallback now covers the long tail: Chrome before 151, Firefox before 155 and Safari before 27 all get the opaque colour, which is why these functions are safe to ship this week rather than next year. Is a translucent chip worth two declarations? For a design system with nine brand colours, yes, easily.
Do the Six From June Still Hold Up?
All six shipped, and one gained a detail worth knowing.
appearance: base-select with ::picker-icon, ::checkmark and <selectedcontent> is in. Anchor positioning follows transform, translate, rotate and scale, and the rename to anchor-valid and anchor-visible is final, with position-anchor now defaulting to normal. :heading is in, still carrying the specificity trap I flagged: it's a pseudo-class, so swapping your six-element reset for it strengthens the rule as well as shortening it. revert-rule, stretch and the subpixel inline layout work all landed as described.
So the June advice stands. If you planned your Safari 27 work off the beta post, you planned it correctly, which is the useful thing to be able to say about a beta.
What Is the Number Nobody Quotes?
844 resolved issues, against 525 announced at WWDC. That means roughly 319 more fixes than the WWDC announcement, and WebKit says 66 of the release's fixes are in SVG.
For interface work a release's fix count tells you more about your next six months than its feature count does, because features you adopt when you choose to and fixes decide whether the thing you already shipped behaves the same in two browsers.
That's the opinion, and you can disagree with it: I find the number more persuasive than the feature list. WebKit calls the jump "60% more" than the 525 it announced at WWDC. That's a quality release wearing a feature release's headline.
What Is Safari MCP?
Safari 27 ships an MCP server so coding agents can drive the browser. I'm noting it rather than recommending it, for the same reason I said to prototype rather than ship the <model> element: it's single-vendor, and the interesting question is whether the protocol survives contact with a second engine. Worth an afternoon. Not worth a workflow yet.
Summary
Safari 27.0 landed on September 17, 2026 with 80 features, 22 more than the June beta, and 844 resolved issues against the 525 announced at WWDC. For interface work the delta comes down to four things: overflow-anchor brings scroll anchoring on by default, so content loading above the viewport stops stealing the reader's place; alpha(from var(--brand) / 80%) collapses the relative-colour boilerplate that token systems were drowning in, alongside multi-colour color-mix(), image(<color>) and light-dark() accepting images; progress() gains no-clamp for motion that should extrapolate past its bounds rather than flatten; and :host:has() finally lets a web component style itself from its own children's state. Everything I flagged in the beta shipped unchanged, so the six from June remain the six.
Try our own free generator
Colour is next, and the four additions only make sense against a palette that already works. Build one below: pick four colors, watch them land on a live interface, and read the WCAG contrast before the next section starts rewriting how you'd declare them.