DEVONthink Pro for Mac

What's new in this version:

New:
- DEVONthink supports Apple's Liquid Glass design language on macOS Tahoe and has a new, fitting app icon. We have modified many parts of the user interface and will continue to do so while Apple addresses issues. Many refinements we've made, e.g., to the popovers, apply to all supported system versions.
- The Data > New > Sheet panel lets you name the new sheet.
- An Open button in the Annotations & Reminders > Annotations inspector lets you quickly open the current annotation file.
- The Versions inspector supports manually creating named versions.
- The settings option Files > Markdown: Preview: Smart quotes and dashes tells DEVONthink to render quotes and dashes using their typographically correct variants.
- A warning shows when a database may be open in another user account.
- After updating DEVONthink to the next feature release, a What's New panel shows the top features of the release.
- A hidden preference, `WindowToolbarTitleHidden`, lets you hide a main window's title and gives you more space for toolbar items. Note that you cannot use it with the `WindowToolbarExpanded` hidden preference.
- The `unlinked Wiki links` AppleScript property for records helps detect WikiLinks without a linked item.
- The `HTTP response` class has two new AppleScript properties, `error code` and `error domain`, to aid in troubleshooting.

New (Pro):
- New settings in the Chat assistant let you dynamically change options for the current chat, including the Search scope, Usage, and Documents mode.
- The Chat assistant now saves up to 50 chats, defined via the option Max. Recent Chats in the AI > Chat settings. DEVONthink automatically names previous chats and lets you access or remove them via the button in both the Chat inspector and popover. Up to ten of today's chats are listed at the top with previous chats in date-specific submenus. Selecting a previous chat will attempt to reveal or open the associated document or selection with the previously used document mode.
- You can open the Chat popover via the keyboard shortcut ⌃⌘A that we added to Tools > Chat.
- The new menu command Tools > Graph opens the Graph inspector in a floating, resizable popover. This popover will be restored on launch and with workspaces.
- Separate AI > Search settings let you control where you want AI to search. This also includes new Web & Wikipedia settings to search via DEVONthink (slower but more up to date), Perplexity, or Exa (faster, but using indexed content).
- The AI > Research Report template is now also available in a PDF variant.
- Support for DeepSeek 3.2 and 3.2 (Speciale) via OpenRouter.
- Support for the image generation model GPT-Image-1-Mini, accessible in the AI > Image Generation settings and the Generate Image panel.
- Support for Google Nano Banana Pro, generating images in 2K and 4K. Note that it generates images more slowly and is more expensive than Google Nano Banana. Also, the Mixture-of-Agents - Images template uses Google Nano Banana Pro, if available.

Improved:
- Main and document windows now use modern APIs on macOS Sequoia and later, providing animated toggling of the sidebars and inspectors. You can also collapse these panes by dragging the separator to the window's edge.
- On macOS Sequoia and later, the Navigate sidebar is now spring-loaded and will automatically pop open if you drag an item to the left side of a main window.
- The Format Bar is available in fullscreen window mode on macOS 15 and later. You show or hide it with Format > Show/Hide Format Bar.
- The selected row in the Move To and Insert Link popovers has better contrast when they don't have focus.
- DEVONthink more consistently positions popovers relative to the current window.
- You can make the group selector larger than before.
- The appearance of informational text is more consistent throughout the interface.
- On macOS Tahoe, PDF documents now display in dark mode when General > Appearance > Use dark background for documents is enabled.
- When converting a rich text document to PDF, DEVONthink checks if the source document only contains a single PDF attachment and no useful text, then uses the found PDF. This also applies to email conversions in the Pro and Server editions.
- New entries made via Data > Add To > Table of Contents can use the bounds of a selection instead of just the page.
- The Table of Contents inspector is more compatible with poorly constructed HTML content, e.g., created by some AI models.
- The "Remind Me" submenu in the Data and context menus is now Annotate & Remind. You can still set a due date on a selected item but you can also create a new annotation document via the template of your choice. The annotation will open in a new document window or be displayed in the Annotations & Reminders > Annotations inspector, if it's open.
- The Insert Quote command of the Info > Annotations & Reminders inspector supports also Markdown documents.
- In the AI > See Also inspector, the keyboard shortcut for the Move To command is only shown when the ⌃ Control key is pressed.
- The context menu of the item list of main windows now includes a Reveal command to reveal replicants, duplicates, and an item's annotation document.
- The command Apply Rule is now available in the context menu for smart rules added to the Favorites.
- The document date placeholders submenus are now under the Document options.
- Using Data > Mark to lock or unlock a group applies the same to its children. This does not apply when locking an item via scripting or smart rules.
- Drag-and-drop or copy-and-paste is now more compatible with third-party editors.
- DEVONthink no longer shows a warning dialog after it crashed or was quit forcefully.
- Automation prompts to allow DEVONthink to control the Finder no longer self-dismiss.
- The `compare` AppleScript command now only returns a maximum of 40 results, consistent with the See Also & Classify inspector.
- You can now control how long the `download URL` and `download markup/JSON from` AppleScript commands run by using a `with timeout of x seconds… end timeout` block.

Improved (Pro):
- Bookmarks are now supported in many AI-driven functions, like labeling, rating, and tagging via the Data menu, summarization, as well as smart rules, batch processing, and AI placeholders.
- When using an AI model supporting tool calls, e.g., the ChatGPT models, the Chat assistant can temporarily highlight words or sentences in the currently viewed document, similar to the Search inspector.
- The Chat assistant can now create comma or semicolon-delimited sheets (`.csv`), instead of just tab-delimited (`.tsv`) ones.
- The "Research Overview" template is now AI > Research Report - HTML.
- The AI > Research Report and AI > The Devonian Times templates use updated models: Gemini 3 Pro or Flash Preview, GPT 5.2, or GPT 5 Mini.
- The Help assistant now first tries to determine the most relevant help pages. This can drastically reduce the number of tokens used, up to 90%, and allows use of newer or more expensive (and more capable) models.
- Updated Google Nano Banana from preview to final release.
- Aspect Ratio replaces the "Size" option in the AI > Image Generation settings and Generate Image panel. Also, the `aspect ratio` AppleScript parameter replaces the "size" parameter of the `download image for prompt` command. Supported sizes are `1:1`, `16:9 (9:16)`, and `4:3 (3:4)`. This setting as well as which generative model is used and, optionally, the quality setting, control the size of the resulting image.
- The Chat assistant and AI actions like tagging or rating, better handle errors. The assistant also more reliably clears chats. DEVONthink better filters AI-generated error messages like those from placeholders in smart rules and batch processing, and also handles OpenRouter errors better, e.g., when a provider is overloaded.
- Updated GPT 5.1 to GPT 5.2, natively and via OpenRouter. Also, GPT 5 Pro, 5.1 Codex, and 5.1 Codex Mini are available in the Chat assistant via OpenRouter, the Codex and Codex Mini models in the Script assistant.
- Updated GPT-Image-1 to GPT-Image-1.5. The Data > New From Template > AI > Mixture-of-Agents template support the new model.
- Updated Anthropic Claude 4.1 Opus to 4.5 and Haiku to version 4.5, natively and via OpenRouter. Claude 4.5 Opus is less expensive, a new parameter reduces response time and costs. Claude 4.5 Haiku supports reasoning, more output tokens, and has knowledge up to July 2025.
- Updated Gemini 2.5 Pro to Gemini 3 Pro and Gemini 2 Flash to Gemini 3 Flash (Preview), natively and via OpenRouter. Also, responses to vision requests are quicker, cheaper, and support a vision limit of 200 items, useful for video, multi-page PDF documents, and multiple selected images.
- Updated support for Mistral Large 3, including via OpenRouter. Also Mistral's Devstral Medium model is now Devstral 2, with a context window of 256,000 tokens and supported in the Script assistant.
- Mistral and local models have improved tool call support.
- Ollama, LM Studio, and OpenAI (Compatible) models now require a minimum context window of 4,096 tokens. We recommend 16k (16,384), when possible.
- Depending on the usage setting, LM Studio can process images with a maximum dimension up to 2,048 pixels.
- DEVONthink examines and filters OpenAI (Compatible) models for capabilities, e.g., vision or tool call support, with servers not reporting this information on their own. Note that the context window still needs to be specified. This provider is also more compatible with endpoints hosting models like GPT 5.x or the O-models.
- Updated further models usable via OpenRouter: Grok 4 Fast to Grok 4.1 Fast, Nova Lite 1.0 to Nova 2 Lite, and zAI's GLM to version 4.7.
- Better support for the Qwen3 models.
- OpenRouter models now prioritize price (Cheapest) or thoroughness (Best) for models having multiple providers, depending on the AI > Chat > Usage setting.
- The `Chat - Query` smart action works more reliably with low-tier models like GPT 5 Nano or Mini.
- The `get chat response for message` AppleScript command has a better description.
- The `as` parameter for the `get chat response for message` AppleScript command accepts `PDF` to return a paginated PDF document of the response.
- The appearance of the OCR settings is more consistent with other settings panes.

Improved (Sync):
- Bonjour passwords stored in Keychain Access now use the host's ID as the host name of a device can sometimes change when switching networks.
- Fixed:
- Ordered lists in Markdown didn't render properly if there was no space after the dot. Fixed.
- Inline MathJax equations didn't always render properly in Asian language documents. Fixed.
- Web pages using the Consent framework crashed WebKit. Fixed.
- Converting emails (Pro and Server editions) and rich text documents to PDF documents could stall the app. Fixed.
- Converting HTML documents didn't convert code blocks correctly. Fixed.
- Capturing a web page within DEVONthink didn't honor the clutter-free setting. Fixed.
- Searching for tags in a toolbar search or via smart groups and rules didn't work correctly with groups excluded from tagging and grouped RSS feeds. Fixed.
- Switching between light and dark modes could cause visual artifacts in labels, ratings, and the Activity pane. Fixed.
- Destination popups, e.g., in the Sorter, didn't display correctly in dark mode if the destination was the root of the Global Inbox or database. Fixed.
- When detached, the Information popover didn't update when switching tabs in the frontmost document window. Fixed.
- The header of tables in various inspectors, e.g., Concordance > List showed artifacts on macOS Tahoe. Fixed.
- For favorited groups, the item count wasn't shown and their unread count didn't match what was reported in the item list. Fixed.
- The internal versions group was neither properly filtered when determining statistics in Database Properties nor was it ignored in toolbar searches. Fixed.
- Deleting a version from the Versions inspector would incorrectly close the current tab or document window. Fixed.
- The Tools > Show Info command closed the popover instead of activating it. Fixed.
- The context menu could display an unnecessary and disabled Put Back command. Fixed.
- The keystroke ⌃↩ on modern versions of macOS didn't work properly in Icon view and could also change text selections. Fixed.
- Pressing the ⎋ Escape key with the database authentication panel open caused an issue. Fixed.
- The Sorter didn't always appear in the menubar on macOS Tahoe. Fixed.
- The `Apply Rule` smart action didn't work properly if the applied rule ran asynchronous actions, like chat suggestions. Fixed.

Fixed (Pro):
- In rare cases, while in List view with a specific custom metadata field shown, disabling that field on another device and synchronizing could crash the app. Fixed.
- The Tools > Chat command didn't close an already open popover. Fixed.
- The prompt field didn't have focus when the Chat popover opened. Fixed.
- The Model popup in the Script assistant was too narrow for many OpenRouter models. Fixed.
- If the same model was accessible via a local and a remote Ollama installation, the local model was chosen in the Model popup in the Chat assistant or AI-driven smart actions, e.g., `Chat - Query`. Fixed.
- Embedded models in Ollama were not filtered from the model list. Fixed.
- The Chat toolbar item wasn't disabled when a Chat popover was already open. Fixed.
- Markdown blocks in chat messages were sometimes stripped out during chat or when saving the conversation. Fixed.
- A document containing text provided to the smart actions `Chat - Query` and `Chat - Continue if` wasn't handled correctly if the document mode was set to Without Document or Document (Vision). Fixed.
- The Data > New > Script window had no title. Fixed.
- The Quality setting in the Data > New > Generate Image panel wasn't used. Fixed.
- Using Data > Move to Trash could have caused glitches if the icon view of the scanner interface was active. Fixed.
- Tags were not applied to a document when scanning to a binder. Fixed.
- The AI commands in the Rating, Label, and Tags submenus of the Data menu incorrectly showed a star when running the Pro or Server editions. Fixed.
- Changing the type of custom metadata fields, e.g., from Boolean or a number to a string, then editing a smart rule using these fields, could crash the app. Fixed.

Removed (Pro):
- Removed tool calls for various AI functions from the Chat assistant: for compiling scripts, support prompts, and for showing a toolbar search. All these functions were superseded by the Search, Help, and Script assistants. These are less expensive, have shorter response times, and are more reliable.
- Removed Claude Sonnet 4 and 3.5 Haiku, both replaced by the 4.5 models.
- Removed the deprecated Sonar Reasoning model.
- Removed the outdated Dall-E 3 image generation model.
- Removed Grok 4 Fast from OpenRouter, replaced by Grok 4.1 Fast.
- Removed Devstral Small and Medium, including from OpenRouter, replaced by Devstral 2.
- Removed DeepSeek 3.1 Terminus from OpenRouter.
- Removed the Exacto models from OpenRouter as they weren't as reliable as expected, as well as support for Voxtral, GPT-5-image-1, GPT-5-image-mini-1, o4-mini-deep-research, and o3-deep-research.