What's new

Claude Sonnet 5 joins the model picker for Pro subscribers at twenty credits, with Claude Haiku 4.5 at five and Claude Opus 5 at thirty under "Show all".

Sonnet and Opus offer the low, medium, high and extra-high effort levels, with high as the starting point.

All three reuse the unchanging part of each request between messages, which is what keeps them at the credit prices shown rather than several times those.

Generated pictures keep the detail the subject calls for — scenes, ornament and painterly styles no longer flatten into a few large blocks of colour.

Shading now comes through as a few flat tones per colour, so a drawing has depth without gaining the smooth blends that cannot be traced into shapes.

Nothing assumes a colour scheme on your behalf any more, so asking for an icon gives you a full-colour one and asking for a monochrome icon gives you a single-colour silhouette you can recolour in one edit.

Two instructions that no image generator can act on have been dropped, freeing that space for describing the picture you actually want.

The assistant moves, restyles, duplicates and deletes shapes by identity rather than by line number, so edits no longer depend on how the file is formatted.

Six new tools do that work: query to find out what is in the drawing, then set_text, set_attribute, set_style_rule, insert_element and remove_element to change it.

Recolouring a diagram exported from a design tool now changes the one rule in its style block, rather than trying to set a fill on each shape and appearing to do nothing.

It can look through the drawing to answer questions such as "where is the text?" instead of replying that it cannot find anything.

Text split across several tspans is now recognised as one label, so asking to change wording works on drawings exported from design tools.

Selecting a shape in the preview tells the assistant which element you picked, and the badge names that element instead of showing raw path data.

When a turn runs out of tool calls the assistant says so and offers to carry on, rather than stopping as though it had finished.

Saving a document immediately after opening it no longer erases its owner.

Safari and other WebKit browsers no longer stall on half-loaded pages or hang while saving.

New opens a genuinely new document instead of reusing the previous document id and showing a blank canvas.

The preview sizes drawings measured in percent correctly, and no longer shows a scrollbar for a drawing that fits.

The editor shows your document immediately and swaps in the full code editor once it has arrived, instead of holding the screen on a loading message.

Phones pick their layout before rendering, so they no longer download the desktop code editor at all.

The header no longer jumps as the logo decodes, and the advert sits below the fold on phones rather than pushing the drawing down.

The editor opens even when signing in fails, instead of leaving you on an empty page.

The cookie banner no longer sits on top of dialog buttons.

A public gallery of shared drawings, with a filter over titles and descriptions.

Cloning a gallery entry gives you the drawing and the AI conversation that produced it, as a draft of your own.

Sharing has three states — private, unlisted, and published to the gallery — chosen from a Share menu.

Gallery entries are published under CC0 1.0, so anything you find there is yours to use.

The publish dialog can write a title and description for you with AI, for one credit.

The Info tab says whose document you are looking at.

Every page carries its own title, description and canonical link, and the static pages are prerendered so they read correctly when shared or crawled.

The gpt-5.6 tiers are available, alongside Kimi-K2.6 on the free tier.

The picker is grouped into free and pro tiers and curated down to one model per price point.

Every model now reasons at high effort by default, which is the setting most edits benefit from.

Declining an image modification no longer swallows the follow-up work the assistant had queued up.

Conversations continue after a reload, and on another device.

Drawings you never saved appear as drafts on the Files page instead of disappearing.

Guests can write a prompt before signing in; the prompt is sent for you once you are in.

Undo is derived from the drawing attached to each message, so the history survives a reload.

Conversations made before this change are carried up from the browser the next time you open the document.

Deleting a document deletes its conversation with it.

Signing in is required before using the AI assistant or buying credits.

A warning appears while your credit balance is running low, rather than at the moment a request fails.

Every place an SVG is put on the page now goes through one sanitizer, closing the gaps where a crafted drawing could run script.

Signing in redirects rather than opening a popup, which the stricter browsers were blocking.

A rebuilt interface with a resizable live preview, a light and a dark theme, and a layout that works on a phone.

An AI assistant that edits your drawing from a plain-language instruction and shows the change as a diff before you keep it.

Image generation from a text prompt, vectorized in the browser into editable SVG paths.

Search across open-source icon sets and insert an icon as inline paths, with no external dependency.

Cloud storage for your drawings, with thumbnails and shareable links.

Autocomplete and hover documentation come from the W3C SVG specification, so they cover the whole element and attribute set.

Click an element in the preview to jump to its code, with a bounding box drawn around it.