Leo 0.5.0 release notes
ATR-2 in preview
Despite being relatively early in training, our new model, ATR-2, outperformed all 15 other models we tested internally, including the latest and next-best general models from Google (Gemini 4 Carbon), Anthropic (Fable 5.2), OpenAI (GPT-6.1 Terra), and Transkribus (Text Titan II). It made fewer word errors than any other model, and its lead widened when we tested for original letterforms and abbreviation signs.
We are currently preparing a second round of training for ATR-2, with cleaner and more diverse data. We are also working on the areas where it still makes errors, including margin notes and columns that are out of reading order, text that is sometimes omitted, and very long, dense images that it can struggle to track.
Although ATR-2 is still in development, since it already outperforms ATR-1 and other models by a wide margin, we are making it available in preview immediately. After its official release, ATR-2 will not only provide more accurate transcriptions, but will also output bounding boxes (region information about each line), which allow “spotlighting”: viewing a line in the image and in the transcription side by side in the app.
Major changes
- Import by link. Paste a link to an image, a PDF or a IIIF manifest in the Upload window and Leo will show you a preview of what it will import before you confirm. You can also import a section or range of images from a IIIF manifest, or up to 20 items from a collection.
- Comments and highlights. You can select words to highlight them in gold, green, blue, purple or rose, or to add a comment in the margin beside the text. On shared documents, people can reply to a comment, resolve it and mention each other by name.
- Share with people, by role. Invite people to a folder or document by email as a Viewer, a Contributor or an Editor. You can change the permissions of each role in the Share window. Anything shared with you shows up under Shared with you in Folders.
- Version history. Version history, under the ··· menu under a text, lists all saved versions of a text, newest first. Give versions that matter to you a name, open older versions to read them, and restore them, either into the current tab or into a new tab. If a new transcription run replaces a text that you have edited, your version will be preserved in Version history.
- Activity. The bell in the top bar now shows you what needs your attention and what has changed in your shared folders. Any tasks currently running will also appear here, and the Tasks panel is no longer available separately.
- Link anything to anything. Type [[ in any text to create a link to a document, image, note or passage. The link will always show the target’s current name, and if you hover over the link, you’ll see a preview.
Working together
- Suggested edits. Changes made by a Contributor will be offered as suggestions that you can accept or reject, and a Contributor can suggest a note that you can accept. An Editor can turn Suggest edits on or off.
- Public links are a switch. Your document is private by default. To let anyone with the link view it, turn on Anyone with the link can view. Comments are not visible on public links.
- Notification settings. Settings has a new Notifications page, where you can choose what Leo notifies you about for each shared folder or document. Mentions will always be sent to you by email.
- An Activity page, and Undo in the top bar. Settings › Account has a new Activity page, with the ability to filter by folder and by person, and a Mark all as read button. Next to the bell, there is now an Undo button.
- When two people edit the same text. If a text is changed by someone else while you have it open, a bar tells you who changed it, when, and offers you the options Show theirs or Keep mine. You can also see who else is working on an image.
- Shared work uses your credits. If you transcribe someone else’s document, the transcription will use your credits, and the Transcribe window will tell you. A document can only be in one shared folder at a time.
History and undo
- Compare versions. Compare puts two versions side by side, with the changed letters highlighted and the unaltered lines collapsed. Changes since displays each change from an older version, so you can accept or reject them one at a time, or all together.
- Compare two tabs. Compare with… in a tab’s menu compares the tab with another. Compare has a one-column view, an overview of all changes, and options for what constitutes a change.
- Undo for more than typing. Your actions lists the actions you’ve recently performed, and offers an Undo option next to each one. These include moves to Trash, deleted comments, status changes, restores, and sharing changes. Pressing ⌘Z outside of a text will undo the most recent change to an image, whether you rotated, cropped, replaced, added, renamed, or deleted it.
- Undo follows you. You can still undo after navigating to another image or tab and back.
- Provenance. Version history can colour the text to show who wrote what, so you can distinguish the words of a model from your own.
Links and notes
- Cite a passage. Select some words and choose Cite to quote them in a note, with a link back to the original line. If the text is later changed, the note will let you know and suggest an updated quotation.
- Notes. Annotation tabs are now called Notes. A note can be attached to a single image or the whole document.
- Documents made of notes. A document no longer has to include images. A notes-only document opens in Text, and you can share and export it just like any other document.
Reading and editing
- Image, Split and Text. A single layout menu toggles between displaying just the image, the image with its text docked alongside, or just the text. In Text, a new Aa menu controls the font size, line length, and typeface, from a selection of nine new fonts, several of which are designed for early modern texts. Until you select another, texts will default to Lora, the first of these fonts.
- Two texts side by side. Text has layouts: One column, or Side-by-side, with two texts displayed next to each other.
- Margin notes sit in the margin. Margin notes will now display in a separate margin next to the line to which they belong, and this margin will shrink and grow with the width of the panel. On a phone and in very narrow panels they will appear below that line. To add a margin note, select the text and choose Left margin note or Right margin note.
- Find and replace. Press ⌘F to search for a term across all the images in a tab, and replace it either once or everywhere. Finalized texts are ignored.
- Marks for editors. The selection menu can now insert an expansion, an illegible word or a margin note with a single click, and the Characters palette includes the letters that don’t appear on a keyboard (such as ſ, þ and ȝ). Square brackets, deletions and tables are now more clearly rendered in the text.
- Line by line. Transcriptions display line numbers by default. If a transcription has its lines outlined, hover over a line to highlight its position on the image, or select multiple lines to highlight them all.
- Not saved. If saving fails, Leo will show you Not saved and offer to try again, and it will warn you if you try to leave the page.
- Finalized text is locked. A Finalized transcription is now locked and cannot be edited until you change its status.
- Cleaner pasting. Paste text into a transcription and it comes in stripped of its fonts, headings and lists. Italics, underlining, superscripts, deletions and links are preserved.
- Contents does more. Contents now has its own filter, its own slider to adjust thumbnail size, and an Expand all button. You can also drag and drop an image’s tabs to reorder them.
Search
- Closer matches. If you search for multiple words and there are too many results, Leo suggests searching for the exact phrase. If it finds no results, Leo suggests alternative spellings.
- Folders in search. Search results now show folders that match your search, along with the number of documents in the folder and the parent folder. Selecting a folder from the search menu will open the folder.
Import and export
- Larger images and longer PDFs. Leo now supports images up to 268 megapixels. A PDF import now tracks the number of images as it goes and resumes if the import is interrupted.
- Clearer uploads. Files that Leo can’t read are skipped, and a message will identify them. If a PDF contains no images, Leo will tell you and suggest re-uploading it with Pages selected. If multiple PDFs are being imported into a new document, and the first fails, the remaining PDFs will still be imported.
- Transcribe while images arrive. You can begin a transcription while images are still uploading or importing. Leo prices the run when you click Transcribe and will start each image as it arrives.
- Add to this document. You can add images to an open document by link, from a PDF link or from a IIIF manifest.
- Crop, split and rotate in the background. Crops, splits and rotations now run as background tasks, so they can survive a page reload or closed tab. Batch crop lets you apply a single crop frame to multiple images, and you can fine-tune each image before saving. When cropping multiple images, the document docks to the side of the images, which are displayed in rows, just as they are in the image list. Closing it restores your previous layout.
- Export more, and more precisely. Markdown is now an available export format. You can choose which images and tabs to export, and whether to include comments and highlights.
- A new look for exports. Exports in PDF and Word are typeset in Lora and include Leo’s name and the date of export at the top of the first page, and the title and the page number at the bottom of each page.
- Truer exports. In PDF and Word exports, margin notes are now placed in the margins. In other export formats, they are marked as notes. Tables are preserved when line breaks are off, lists retain their numbering, and links to images in the text open the images from the export. Word exports preserve highlights and heading levels, plain text exports have readable tables, TEI exports are valid, and right-to-left text is displayed correctly.
- Markdown that keeps its shape. Markdown exports now preserve your headings, margin notes and underlining, merged cells in tables, alignment, text size and highlight colours.
- Exports in more scripts. PDF and Word exports now embed fonts, so Greek, Hebrew, Arabic, and early modern characters are printed correctly. Comments can now be exported in all formats, and a completed export can be redownloaded for 7 days.
Models
- A shorter model menu. The model menu now includes only the latest models from each provider, meaning that around 30 older models are no longer available for selection. Existing transcriptions remain unchanged. If your saved model was one that is no longer available, your next run will use ATR-2.
- New rates. Credits per image have changed with the new list: Claude Fable 5.2 is now 2.27 credits, Claude Opus 5.5 is now 1.12, Gemini 4 Carbon is now 0.83, Gemini 4 Argon is now 0.76 and Gemini 4 Hydrogen is now 0.55. A run still costs at least 1 credit.
- Gemini’s thinking counts in transformations. If you transform text or generate metadata with a Gemini model, the tokens it spends thinking now count toward your token limit. Transcriptions are still priced in credits per image, as before.
- Third-party AI switches. Settings › Privacy and cookies now includes four switches for the external AI services that Leo can call: AI-model transcription, text-reading services, transformations and Generate metadata. All four are on by default.
- Choose a model for one run. Transform and Generate metadata now allow you to choose a model for a single run and show you the estimated cost. Generate metadata will now display the result in the same window.
A cleaner, faster Leo
- One look throughout. Buttons, menus, form fields, panel headings, and dialogs have been redrawn to have a consistent size and hover effect, in light and dark mode. Settings has been reorganised, and Appearance has Light, Dark, and Match my system options.
- Faster on large accounts. Accounts with hundreds of folders load their folder lists reliably. On the largest accounts, transformations start in about a second, transcription runs start without timing out, and document thumbnails load sharply. In a long document, thumbnails are ready before you scroll to them.
API changes
- Connect to Claude and ChatGPT. Settings › API has buttons to connect Leo to Claude or to ChatGPT.
- Runs default to ATR-2. Transcriptions that specify no model, or specify a retired one, run on ATR-2.
list_transcription_modelslists the current models only. - ATR-2 is marked as a preview. The API and connected apps refer to ATR-2 as a preview in any place that lists it. During its preview period, all runs use default settings and return the transcribed text without line boxes. A request to set formatting or bounding-box options will be rejected with an explanation of why.
- See tryleo.ai/developers for the complete list of changes.
Minor changes
- Model selection is now contained within the Transcribe, Transform and Generate metadata windows rather than in a floating menu
- The Canvas layout now includes Anchored, so there are now three layouts: Docked, Canvas, and Tiled
- The Images panel is now called Contents, and each image expands to show its tabs, notes, comments and highlights
- The Transcription column on the dashboard now displays the number of transcribed images, e.g. 12 of 40
- Back takes you to the screen you were on before, and each document and each section of Settings is its own step
- Links opened within Leo open in place
- Saving and Saved indicators now appear next to the image as well as the text
- Repeated rotations of an image are now queued
- Progress is now shown when copying or moving documents
- A new document now shows the message Images are being imported while it is being filled from a PDF, link, or manifest
- Exporting to ZIP will name the files according to their full image name, and ZIP export of texts can include highlights
- Plans and billing are now in Settings, and the credits popover shows your total
- Settings › Billing shows the name of the payment method you have saved
- Plan changes preview your charge for today before you confirm
- Settings › API now displays your usage chart, recent requests, and webhook deliveries on a single screen
- Webhooks are disabled when the API key that created them is revoked
- Version history is now a sheet on a phone, and the layout menu now has options for Image or Text
- Touch targets are larger on phones and tablets
Bug fixes
- Multiple images selected or dragged at once go into a single document again
- A transcription with a second model no longer overwrites the existing transcription on the image
- A transcription that completes after Leo has merged its tab into another will now arrive in the merged tab
- Large JPEGs that previously caused long uploads to fail midway can now be imported
- A canceled task now actually cancels, and a stalled task no longer prevents other tasks from running
- A failed PDF import no longer leaves behind an empty document
- Exporting from the document menu now works on a phone
- Deleting a very large document no longer results in an error
- Improved security of the app and the server
More updates to follow in the coming weeks!