📌 Release Overview

• 2026.44

• Unique hosted:October 28, 2026

Executive Summary: highlights of the release

Conduct

• GA Model and effort choice per Conduct space: On by default; admins set models and efforts per space and can let users switch per message.

• Beta Pi Agent: It's a new Conduct harness that runs on models from many providers, including OpenAI, Anthropic, Google, Grok, DeepSeek, Moonshot and Z.ai, and is the only harness that supports Gemini and open models such as Gemma or Mistral. That lets you save on token costs and keep confidential data on your own servers by running self-hosted open models, while Pi itself is lighter and faster than Claude Code and Codex.

Chat Platform

• GA MCP for every company: Out of Beta and always available in chat and admin.

Agentic Table

• GA Multi-assignee dropdown: The assignee column is redesigned to match the rest of the platform, so assigning people looks and works the same everywhere.

• Beta Stop a running table: The Updating... badge holds a Stop button, and the sheet is available again after Stop. Off by default, enabled on request.

Check before upgrading:

• Model and effort choice in Conduct spaces is on by default: companies with Conduct get these controls after upgrading unless it is turned off per company (see Conduct).

• MCP is always on, including for companies that had it turned off (see Chat Platform).

• Deployments that use Virtual Servers must set the new company setting, or the Virtual Server pages disappear from admin (see 2026.44 Infrastructure Changes).

• Every company gets a templates-conduct folder at its Knowledge Base root, and every active user gets a home folder during the upgrade (see Conduct and RAG).

• Integrations using the Public API: model failures now return 429, 503 or 504, a restricted model 403, and short-term memory requests for other users’ chats 404 (see Chat Platform).

• Agentic Table Word exports become private to the person who created them (see Agentic Table).

• Feature flags, settings and operator changes (new, changed and removed) are listed on 2026.44 Infrastructure Changes.

Conduct

🚀 What’s New (Business/End-User Focus)

• Admin GA Model and effort choice per Conduct space: Admins pick each space’s default model and effort on the Agent Harness card, and can let users switch to further models with the efforts they allow.

  

  • Coding Agent settings moved to the harness card’s Advanced Settings. See Unique AI Conduct Space.

    

  • Behavior change: This is now on by default, so companies with Conduct get these controls after upgrading. It can be turned off per company (FEATURE_FLAG_ENABLE_CONDUCT_MODEL_SELECTION_UN_26651).

• User GA Pick a model and effort per message: In spaces that allow it, the model list shows each allowed effort as its own entry, such as GPT-5 (Low) and GPT-5 (High).

• 

  Admin Beta Pi Agent: Pi Agent (based on pi.dev) is a new agent harness in Conduct that isn't tied to a single model provider. Claude Code is built around Anthropic's models and Codex around OpenAI's. Pi runs on closed- and open-sourced models from OpenAI, Anthropic, Gemini, Grok, DeepSeek, Mistral, Moonshot and Z.ai. Open models usually cost less per token, so teams can cut costs on everyday work. Organisations that are not allowed to send confidential data to a provider such as OpenAI or Anthropic can connect Pi to self-hosted models, so the data never leaves their infrastructure. Pi is also lighter and faster than Claude Code and Codex, and independent benchmarks show it can perform better on the same models.

• Admin Templates folder for Conduct: Every company gets a templates-conduct Knowledge Base folder for document templates, readable by all users and not indexed for search. Admins grant write access to whoever maintains templates. See Company Templates for Document Generation Skills.

  • Behavior change: the folder appears at the Knowledge Base root of every company, including companies without Conduct. In spaces whose search is limited to selected folders, search now also covers it.

• Admin More models in Pi, Claude Code and Codex: Claude Haiku 5.5 is selectable in Claude Code and Pi. Pi adds Claude Sonnet 5.5 and Opus 5.5, Grok 4.7, GPT-5.4 Mini, GPT-6 Astra, Luna and Sol, GPT-6.1 Sol, Gemini 3.6 and 3.8 Flash, Gemini 4 Argon, Gemma 4 31B, GLM-5.3 Flash and DeepSeek V4.1 Flash; Codex adds GPT-5.4 Mini.

  • Behavior change: Pi compacts long chats based on each model’s input limit, so compaction can start at a different point than before.

• Admin Install Conduct spaces through the API: The Public API can create and change Conduct spaces with their MCP servers, and shows whether a user has connected each server. See Unique API Version 2026.03.01.

• Admin Skill count per source folder: The space Sources settings show how many skills each folder provides, and the total for additional sources.

🔧 Improvements

• User GPT-5.1 behaviour improvement in Codex:

  • generated file check: Generated files based on a template with placeholders, leftover sample text, missing values, layout problems or missing German, French and Spanish accents go back to the agent to be fixed.

  • Keeping company templates intact: Word and PowerPoint templates keep their cover, headers, footers, layouts, fonts and logos, and a field it cannot fill is removed and named in the answer. See Company Templates for Document Generation Skills.

• User Templates only when needed: Conduct builds ordinary Word and PDF documents right away and asks for a company template only when you name a template, company or brand, or the document is client-facing. Short answers stay in the chat unless you ask for a file.

🐞 Bug Fixes

• User Model switching respects allowed efforts: Switching to a model that cannot run the space’s effort uses the next lower effort the admin allowed. If none fits, the chat asks an admin to update the space.

• User Clearer progress on long turns: A quiet turn says what it is doing and counts down to the real stop time, and a turn stopped at the maximum working time says so in the answer.

• User File delivery on Codex models: Conduct no longer claims a file was saved or sent when no tool did that; it hands over the file and names the destination it could not reach.

• User No extra closing messages: Conduct no longer asks for a closing message after a complete answer, including slow replies.

• User Cancel stops background commands: Cancelling a turn stops leftover commands, so they no longer keep scanning the Knowledge Base.

• User Dashboards open quickly: Charts in HTML dashboards and reports are drawn inside the page, so chart files stay small instead of several megabytes.

• User Several Codex chats at once: Several Codex chats running at the same time no longer stall file uploads and downloads until a restart.

• User Personal skills without Knowledge Base roles: Users without Knowledge Base roles can save personal skills to their home folder and see their personal and the company’s skills. See Skills Management in Unique AI Conduct.

• User Skill pickers sorted: The / menu and Skills in the tools menu list skills alphabetically.

• Admin Skills settings fixes: Skill settings find skills nested up to 11 folders deep, toggles change only when you click them, and the Sources section is evenly spaced.

🔒 Security

• Admin Container images patched: Conduct and assistant images fix published critical and high container vulnerabilities, and the assistants image no longer includes tools that only the sandbox needs.

Chat Platform

🚀 What’s New (Business/End-User Focus)

• User GA MCP for every company: MCP in chat and admin is out of Beta and always available. See The MCP Hub and Use MCP Hub.

  • Behavior change: companies that had MCP turned off now see it in chat and admin.

• Admin Experimental Admin, Theme and Knowledge Upload in German and French: These apps and shared controls such as pagination, file upload and settings follow the interface language saved on the user account, and company copy overrides apply to their text.

• Admin Experimental API usage in Cost Management: A new API Requests tab shows a company’s API requests per period by type, grouped by user, assistant or space type, with CSV export. See Model Usage and Cost Management.

• Admin Experimental Duplicate a space as a draft: Duplicating a space creates an unpublished copy that opens for review and cannot be used in chat until an admin publishes it with Update. See Space Management.

• Admin Experimental Newer speech models for voice input: Voice input can use a newer Azure speech model that detects a language change mid-dictation, or transcribe recordings with Azure MAI-Transcribe, which applies the space’s word list and can drop filler words. See Voice Administration.

• Admin Word replacements for voice input: Admins can list word replacements in a space’s speech settings, so voice input writes domain words the same way every time, for example “Zitadel” instead of “Citadel”.

• User Interactive HTML previews work with your data: Previews that opt in can read chats, messages, Knowledge Base folders and MCP connectors and, while you use them, start chats, send messages and create files where you can write. See HTML Rendering.

• User Select a range in Chat History: Shift-click selects a range of conversations to delete.

• Admin Benchmark alerts by email: Benchmark Space alerts can be sent as email, one digest per space with a link to each alert. See Agent Benchmarking: Evaluations.

• Admin Unique AI defaults through the API: Chat admins can read, replace and clear the Unique AI default configuration through the Public API; changes reach running chats within about 10 seconds. This is not available in the interface.

🔧 Improvements

• User Clear reasons when a model call fails: Chat names the reason, such as a rate limit, quota, content filter, input length, restricted model, daily limit or provider outage. Unexpected errors show a correlation id to quote to support.

  • Behavior change: the Public API now returns 429, 503 or 504 for these failures instead of 422 or 500, and 403 for a restricted model. A daily-limit error no longer shows the spend amount. See Unique API Version 2026.03.01.

• User One top bar on every page: Chat, space, Agentic Table, history, discovery, analytics, cost, access and MCP pages share one responsive top bar on desktop and mobile. For regular chats, the new chat header is off by default.

• User One header for side-panel previews: File previews share one header with Hide chat / Show chat, back and forward, copy, download and a More options menu with Edit and the link graph. The side panel button now reads Show canvas, and Markdown previews hide their frontmatter.

• User Hidden chats stay out of your list: Chats created as hidden, and benchmark runs, no longer appear in the chat list, Recents, Chat History or search; they stay reachable by link.

• User More of the interface in your language: Debug metrics, HTML preview prompts and several error messages follow the selected language.

• User Compact voice recorder: Voice input shows a compact waveform with a timer that fits the message box on every screen size.

• User Debug inspector cards: The chat debug inspector shows each payload in a collapsible card with copy and download.

• User Shared Markdown styles: Markdown outside chat messages uses the same heading, quote, list, task and table styles as chat.

• User Outlook search pages its results: Outlook email search returns short previews and loads more results on request; opening an email still shows all of it. See Outlook Semantic MCP – Features.

  • Behavior change: the first search response no longer includes full email bodies.

• Admin Cost per user: Cost Management breaks an assistant’s cost down per user instead of per model, and drilldown filters name what they filter.

• Admin Redesigned virtual server pages: A virtual server’s tools are managed per connector in one dialog with search, changes save with Update, and leaving with unsaved changes asks first. Cards show connector logos. See Virtual Servers.

• User New Virtual Server connect page: The connect page shows each server’s icon and confirms once every server is connected.

• Admin Virtual Servers switched on per company: Virtual Servers are now turned on with a company setting instead of a feature flag.

  • Behavior change: deployments that used the old flag must set the new setting, or the Virtual Server pages disappear from admin (see 2026.44 Infrastructure Changes).

🐞 Bug Fixes

• User Experimental Scheduled tasks: Runs show their date and time again, Run now explains a missing space or chat, a task card shows its current space or “Space unavailable”, a task turned back on stays on, and the page lists only tasks you set up. See Scheduled Tasks.

• User Generated files keep their versions: Earlier answers keep showing the file version they were written with, older versions keep their own content, and previews of an older version show progress while they load.

• User HTML preview fixes: Previews built with bundled libraries render, folder lists show subfolders, and creating a chat from a preview on an assistant’s page works.

• User Layout fixes: Long chat titles truncate, the lock icon of a chat without access sits next to the title with a fully visible tooltip, selected tools stay beside the tools menu on desktop, and Manage Prompts fits smaller screens.

• User Chat controls after editing a space: Chat controls keep working after editing or deleting a customized space, and custom space instructions stay visible on narrow screens.

• User Chat History uses the full width: Opening Chat History from a chat closes the side panel instead of leaving it empty.

• User Skills in Settings: Skills stays in Settings when Context Memory is off, and a cut-off skill path shows in full on hover.

• User Reference names: Reference chips from MCP connectors show the connector name, and cards for externally stored files show the file name.

• User Safari follows streaming answers: Safari keeps a long streaming answer in view, including approval forms shown during it.

• User Voice recording keeps the whole dictation: Recording mode keeps the whole dictation instead of only the last 15 seconds.

• User Spend badge: The user menu shows today’s spend and the remaining daily budget instead of the total limit.

• User Document translation file name: Document translation chats show the uploaded file name instead of an internal id.

• User Visual polish: Resize handles show on standard-resolution displays, scrollbars are lighter, menus show delete actions in neutral styling, and desktop tabs show a hover state.

• User Outlook tools more stable: The Outlook connector releases idle sessions, so Outlook tool calls in chat are interrupted less often.

• Admin MCP admin fixes: Activity records of tool calls from a space show the space, and closing the Limit to spaces dialog no longer freezes the page.

• Admin API fixes: Embeddings requests with a timeout are accepted, and messages and streamed responses report cached and reasoning tokens.

• Admin Client insights counts: Prompts-per-assistant counts include all user prompts again.

• Admin AI Module Template disclaimer legible: The disclaimer text in the Save Changes tooltip is readable again.

🔒 Security

• Admin Short-term memory limited to your chats: The API reads and writes short-term memory only for chats the caller owns or Agentic Table chats they belong to.

  • Behavior change: requests for another user’s chat, or for a deleted chat or message, now return 404.

• Admin Container images patched: Platform services and frontends fix published critical, high and medium vulnerabilities, including a critical Next.js issue.

RAG

🚀 What’s New (Business/End-User Focus)

• User Beta Home folder for every user: Every active user gets a home folder at /home/<UserName> during the upgrade, and new users when their account is created; service users do not get one. See Knowledge Base for End Users.

• Admin Search scoring options in the API: The search API can choose how combined searches merge vector and keyword results and can return relevance scores. See Unique API Version 2026.03.01.

🔧 Improvements

• User Deleted files stay restorable: Deleted versioned files can be restored for the configured file retention period, then are removed. Files deleted earlier get the same period, counted from this release.

• User Experimental Wider file preview in column view: Widening the file preview narrows the folder columns, and the preview can take up to half of the Knowledge Base area.

• User Clearer New Folder dialog: Private and Shared are two explicit options, each with its description.

• User Consistent Knowledge Base header: The Knowledge Base uses the shared responsive top bar and keeps its size and centered search on desktop.

• User Older file versions: Deleting a file also removes its archived versions, and previews of archived Office versions are created on request.

🐞 Bug Fixes

• User Folder-path searches no longer time out: Searches that filter by folder path on large knowledge bases complete, including on existing knowledge bases.

• User Retry when emptying a folder fails: A Retry button deletes the remaining files instead of a misleading error.

• User Office uploads keep previews: New versions of Office files keep their PDF preview, and the previous version’s preview is archived.

• User Uploads with required metadata finish: PDF and Office uploads finish ingesting where required document metadata is configured.

🔒 Security

• User Beta SAgent folders in home folders stay private: Folders that agents create in a user’s home, such as memory and personal skills, and their files can no longer be shared, moved out or renamed. Sharing a single file now needs Manage access. See Knowledge Base for Admins.

  • Behavior change: on upgrade this applies to every existing folder in a home, including folders users created by hand.

• User Paths hide other people’s homes: A folder path no longer shows another person’s home folder.

  • Behavior change: this applies to every company, with no setting to turn it off.

• User Malware results stay with copies: Archived and restored copies keep a file’s malware scan result, so a flagged file stays blocked.

• Admin Container images patched: Ingestion images fix published high and medium vulnerabilities.

Agentic Table

🚀 What’s New (Business/End-User Focus)

• User Collaborators work with source files: People with write access to a sheet, including those it is shared with, can upload and remove its source files; readers can view and search them. Previously only the owner could.

• User Beta Stop a running table: While a table runs, the Updating... badge holds a Stop button. Stopping the run makes the sheet available again, including when processing was interrupted. It is off by default and enabled on request.

• Admin List sheets through the API: The Public API lists the sheets a caller can reach in a space, with their access per sheet. See Unique API Version 2026.03.01.

🔧 Improvements

• User GA Multi-assignee dropdown for everyone: The assignee column is redesigned to match the rest of the platform: stacked initials, a name on hover, and a +N count. Every sheet shows it.

• User Can answer preselected when sharing: Sharing a sheet preselects Can answer for new collaborators, so they see only the rows assigned to them. Existing members keep their roles.

• User Citations follow access: Saving a cell keeps a citation only when the person saving can open that source or it is already cited where they can see it.

• User Experimental Bulk add for editors: Users with Can edit can bulk-add verified rows to the Answer Library when they can edit that library. See Answer Library.

• User Experimental Typed questions get metadata tags: Questions typed in or added through the API get the same metadata tags as imported questions.

• User Re-run from the bulk actions bar: With bulk actions on, re-run selected rows from the bulk actions bar; the sheet-level Re-run button is hidden.

🐞 Bug Fixes

• User Experimental Edit with AI: Edit with AI appears on the Answers column with the default RFP configuration, keeps the original sources, shows a source before you accept, and updates References when you accept. See Edit with AI.

• User Sheets recover after interruptions: Question uploads no longer hang, and a brief network error no longer leaves a sheet on Updating.

• User Add to Library: Editors with write access to the library can save answers, a missing permission is named, and adding twice within a minute keeps every entry.

• User Word exports are private: Each person sees only their own exports, and an export by a Can answer collaborator contains only their rows.

  • Behavior change: a collaborator can no longer download a report the owner exported, and the owner cannot download theirs.

• User Typed questions land at the end: A question typed into an empty row is added at the end of the sheet; if a filter hides it, a message offers to clear the filter.

• User Sheet list keeps its place: Page, sort and search stay when you open a sheet and go back, and after a refresh.

• User Tables in cells scroll: Wide and long tables inside a cell scroll, with each header on one line.

• User Smaller fixes: Review Status is disabled on unsaved rows, the update time sits above the headers, new sheets accept questions right away, and columns using chat citations run again.

🔒 Security

• Admin Container images patched: The Agentic Table image no longer ships curl, which removes published vulnerabilities.

Unique AI

🔧 Improvements

• User Content filter triggers: When the content filter is triggered, a clear message tells the user to rephrase the prompt instead of the generic error.

🐞 Bug Fixes

• User External file names appear in chat references: Chat now shows the file name (in the references) for externally stored files, instead of the full storage URL. Web-search citations keep their current labels, and opening a reference still takes you to the external file.

• User Cached and reasoning tokens: Message telemetry in the debug info now includes the cached input and reasoning token counts that the model provider returns.

🔒 Security

• Admin Container images patched: Assistant images fix published high vulnerabilities in GraphQL and image-processing libraries.

Language Model

Admin Claude Haiku 5.5

Claude Haiku 5.5 is available in Unique AI and in the Conduct Claude Code and Pi model lists, via Anthropic and Vertex AI:

• litellm:anthropic-claude-haiku-5-5

• litellm:vertex-claude-haiku-5-5

Admin Gemini 4 Argon

Gemini 4 Argon is available in Unique AI via Google and Vertex AI, and in the Conduct Pi model list:

• litellm:gemini-4-argon

• litellm:vertex-gemini-4-argon

Web Search

🚀 What’s New (Business/End-User Focus)

• Admin Customized API web search through the Search Proxy: Web searches that use a Customized API can go through the Search Proxy, with the same URL safety checks, size limits and corporate egress proxy as the other search engines. It is turned on per company, and operators can approve Customized API endpoints on private network addresses (see 2026.44 Infrastructure Changes).

🐞 Bug Fixes

• User More reliable Grounding with Bing sources: Source links and titles in Grounding with Bing results now come directly from Bing’s citations.

  • Behavior change: the fallback parsing Language model setting of Grounding with Bing is no longer used and has been removed; existing configurations still load.

• User Improved URL propagation for non-Google engines with Search v2: The model now gets each search result’s URL reliably, so Search v2 spaces using engines that don’t require crawling on every result, such as Brave, can now more reliable crawl returned URLs. Google spaces are not affected.

🔒 Security

• GA Web Search Proxy: Web search traffic now leaves from one dedicated service, separate from the assistants. Legacy paths still work until the service has been deployed and configured.

• Admin Bing and Vertex AI use the corporate egress proxy: In deployments with a corporate egress proxy, Grounding with Bing and Vertex AI searches also go through it, like the other search engines.

• Admin Turning off URL safety checks needs a corporate proxy: URL safety checks can be turned off only when all outbound traffic goes through a corporate proxy.

  • Behavior change: a Search Proxy with the checks turned off and no such proxy does not start