📌 Release Overview

• 2026.42

• Unique hosted: October 14, 2026

Executive Summary: highlights of the release

Conduct

• Experimental GPT-5.1 limits banner: Conduct chats on GPT-5.1 warn about the model’s limits.

Chat Platform

• GA Pinned chats: Pin a chat to keep it at the top of your chat list.

• GA Shared Chats: On by default for every company; each space still needs its Shared chat toggle.

• Experimental Embedded App spaces: A new space type shows an external web app inside Unique.

• Experimental Personal side-panel files and your own dashboard: A space can open a file from each user’s home folder, such as a personal Remy dashboard.

• Experimental Scheduled tasks: Tasks run on time again, keep their local time, check eligibility before every run and survive a deleted space.

RAG

• GA New Knowledge Base interface: On by default, with new list, search and column views.

• Experimental Open a file straight to the cited pages: Knowledge Base links open the preview with the cited pages highlighted.

• Experimental Extracted metadata searchable from the first ingestion: No re-ingestion needed.

• Experimental View access for chat-only users: Chat-only users can be given view access to a Knowledge Base folder.

Agentic Table

• Experimental Add verified rows to the Answer Library in bulk: Sheet managers and owners add selected rows in one step.

• Experimental Group named in the turn-on confirmation: Shows the group a recipient was assigned through.

• Experimental Select all search results in a column filter: Selects only the matching values.

Unique AI

• GA Context Memory: Available to all companies; space administrators switch it on per space.

Check before upgrading:

• Customers who ingest Microsoft Purview protected files must change one Entra permission and set one variable on the ingestion worker (see 2026.42 Infrastructure Changes).

• Companies that have not set the New Knowledge Base interface flag switch to the new interface after upgrading (see RAG).

• Feature flags, settings and operator changes (new, changed and removed) are listed on 2026.42 Infrastructure Changes.

Conduct

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

• User Updated hallucination check flags contradictions: An answer is marked red when a statement contradicts a cited source. An unfinished check shows as not checked.

• Admin New model choices:

  • Claude Code: Claude Opus 5.5, Claude Sonnet 5.5

  • Codex: GPT-6.1 Sol

• User Experimental GPT-5.1 limits banner: Conduct chats on GPT-5.1 warn that the model has limitations with document generation and longer tasks. See GPT-5.1 in Unique AI Conduct. Banner is by default on and can be deactivated with the feature flag FEATURE_FLAG_GPT_51_LIMITS_BANNER_UN_26126.

• User Personal skills now live in your home folder: Skills you create are saved in your own home folder, which only you can open. A skill saved in one space loads only in that space. Settings → Skills lists these with the Personal tag. Existing personal skills were not migrated: they stay available in their current folders, they are not marked Personal in the skills list, and access still follows the permissions on that folder. See user-facing and admin documentation.

  • Note: Creating personal skills currently requires the Knowledge Base Write role. Users with only Chat Basic (chat.chat.basic) cannot save new skills yet. This functionality will become available with the next release.

🔧 Improvements

• User New default hallucination judge model: The judge defaults to GPT-6.1 Sol where available, otherwise to GPT-5.1 at medium reasoning effort. See Hallucination Check: judge model and reasoning effort comparison. Spaces that were using another judge, such as GPT-4o, are automatically migrated to the new default model.

• User GPT-5.1 improvement: Shorter, sharper skills for file PDF and Word generation. Company Word templates are filled in place, keeping their cover, headers and styles.

🐞 Bug Fixes

• User Hallucination check reads the conversation and never blocks it: The check considers earlier messages, so correct follow-ups are no longer flagged, and a slow check no longer locks the chat.

• Admin Groundedness check reads every cited source: Answers citing many emails, Slack or Teams messages are now checked against every cited source, not only the first few.

• User One citation per item for large tool results: Large tool results, such as a Jira search, expand into one citation per item.

• User Long turns keep running: Conduct no longer stops turns that are still working or waiting on a sub-agent, and a turn with no first response restarts once before failing.

• Admin Stuck sub-agent turns stop sooner: A turn stuck on a sub-agent that never finishes stops after 30 minutes without progress instead of one hour.

• User Timeline shows the right steps in order: Thinking steps, tool cards and forms stay in order, and tools that run at the same time each show their own result.

• User Long connected-space calls show as running: Calls longer than 150 seconds no longer show as failed while still running.

• User More faithful documents: Conduct asks where a named company template is instead of building a look-alike. See template documentation.

• User Scripts in skills: Skills saved from a chat can upload their scripts, e.g. .py files, and to the knowledge base.

• User Internal and Web Search exclusivity respected: Spaces with Internal Search no longer use Web Search when exclusivity is enforced via the flag FEATURE_FLAG_ENFORCE_INTERNAL_SEARCH_WEB_SEARCH_EXCLUSIVITY_UN_2297. This flag is already used in UniqueAI chat.

• Admin Coding Agent dialog scrolling: The Coding Agent dialog (e.g. thinking blocks) no longer scrolls into empty space.

• Admin Effort levels match each model: Each Claude and Codex model offers only the effort levels it supports, and Claude now offers Extra high.

🔒 Security

• Admin Retention removes all Conduct chat data: Data retention now also removes elicitation and short-term memory content of expired and deleted chats, including chats that expired earlier.

• User Expired chats leave no fork details: Fork attachment names and progress details are removed when a chat expires.

• Admin Container images patched: Conduct sandbox images fix published critical and high container vulnerabilities.

Chat Platform

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

• User GA Pin chats: Pin a chat from its menu in the sidebar to keep it at the top of your chat list. On by default.

• Admin Experimental Embedded App spaces: A new space type shows an external web app inside Unique, and space instructions can open allowed sites in the side panel. Operators list the allowed sites (see 2026.42 Infrastructure Changes), and the app handles its own sign-in.

• Admin Experimental Personal side-panel files: A space can open a file from each user’s home folder in the side panel, such as $HOME/dashboard.html. If the file is missing, the shared file opens instead.

• User Experimental Your own dashboard in the side panel: A setup page copies a blank Remy dashboard into your home folder, and a Show side panel button in the chat header reopens it. See Extending Remy - skills, actions, pages, sources.

• User Interactive HTML previews do more: Buttons in an HTML preview can place a suggested prompt in the chat input for you to edit before sending, and previews can load Knowledge Base files by path. See HTML Rendering.

• Admin Experimental Scheduled Task ID in the interactions report: The detailed user interactions report shows which scheduled task sent each prompt.

• Admin Choose the model per message through the API: API callers can pick the model for a single message, following the same rules as model switching in chat. See Unique API Version 2026.03.01.

• Admin Message logs in bulk through the API: The API lists tool calls and skill loads for up to 200 messages at once, limited to chats the caller owns.

• User Richer message telemetry: Chat owners can see run statistics such as time to first token, cost and token usage per model without debug access.

🔧 Improvements

• User GA Shared Chats generally available: Chat sharing is on by default for every company; each space still needs its Shared chat toggle. Companies that already chose their own sharing settings keep them.

• User New answer formatting and HTML rendering for every chat: The new answer layout and HTML rendering are now on in every chat. See HTML Rendering.

  • Behavior change: organizations that used the legacy formatting or had HTML rendering disabled now get the new layout and HTML previews.

• User Experimental Scheduled task runs appear as a step: A scheduled task’s prompt shows as a “Ran scheduled task” step on the answer instead of as a message you typed.

• User Experimental Scheduled tasks keep their local time: Tasks run in the time zone they were saved in, also across daylight saving changes, and are always shown in your own time. Existing tasks switch to your time zone the next time you save them.

• User Experimental Scheduled tasks run at most once per slot: A run that cannot start within 5 minutes of its time is skipped, and a failed run waits for the next scheduled time.

• User Experimental Scheduled tasks survive a deleted space: Deleting a space blocks its scheduled tasks instead of deleting them, until the owner picks another space. A task skipped five times in a row is switched off.

  • Behavior change: tasks returned by the API can have an empty assistantId when their space was deleted.

• User Better German and French wording: “Space” and, in French, “Chat” are now translated, as are the message debug panel labels.

• Admin Virtual Server administration per server: MANAGE access on a Virtual Server now grants its admin page, including delegation and deletion. Organisation-wide administration needs the new role virtualservers.admin.manage or Chat Admin. See Virtual Servers.

  • Behavior change: the role mcp.admin.virtualservers is removed.

• Admin Refreshed MCP admin pages: Connector and Virtual Server pages have a new header layout and can be sorted by most recent or by name. See The MCP Hub.

• Admin Billed tokens in message debug info: Message debug info shows the billed tokens per model next to the cost, for users who can read it.

• User Spaces and messages API on the new Unique API: The spaces and messages routes of the Public API now run on the new Unique API; the base URL is unchanged and no client action is needed. Details on 2026.42 Unique API / SDK Changes.

  • Behavior change: on these routes, calls without the required role return 403, invalid payloads return 400 and internal errors return 422.

• Admin Encrypted Redis connections on any port: Operators can enable TLS for a standalone Redis on any port. Settings on 2026.42 Infrastructure Changes.

🐞 Bug Fixes

• User The interface stays in your language: Sign-in screens, notices, error messages, side panel titles and the default greeting now follow the selected interface language. Swiss German uses Swiss spellings.

• Admin Translatable sign-in errors: Sign-in and registration errors carry a stable error code, so they show in the user’s language.

• User Experimental Scheduled tasks run on schedule again: Tasks run on time instead of failing every run and switching themselves off. A task that was switched off automatically explains why.

• User Chat owners can share again: Chat owners without admin rights can search for recipients and copy share links again.

• User Side panel and navigation keep their layout: Opening a large side panel keeps the navigation open on desktop, and the back arrow in the HTML side panel returns to the previous page.

• User Dashboards behave predictably: A fullscreen dashboard returns to split view when it sends a prompt to chat, and HTML previews show the latest version of a replaced file.

• User Mobile and touch fixes: On mobile, Enter adds a new line instead of sending, and on iPhone the header and send button stay visible while the keyboard is open. Rename, delete and other hover controls now show on touchscreens.

• User PDF previews open at 100%: Chat PDF previews open at 100% zoom.

  • Behavior change: Page fit is no longer selected automatically but is still available in the zoom controls.

• User Visual polish: Scrollbars stay hidden until you hover, dialogs dim the page less, and tooltips wrap more cleanly.

• User Live updates recover after an outage: Live chat updates recover within seconds after a cache failover, instead of staying silent until a restart.

• User Connecting MCP servers is smoother: Connecting no longer fails with “OAuth session expired” while a tool call runs, and chats list the MCP servers used by sub-agents so you can connect them. See Use MCP Hub.

• Admin Virtual Servers only offer callable tools: A Virtual Server can only be enabled or published when at least one of its tools can be called. Tools that need a disabled context namespace are locked.

• Admin Connector refresh and badges: Refreshing a connector whose server offers no tools marks those tools Unavailable instead of failing.

• Admin Space and access settings: Custom Spaces keep their Tool Definition parameters on save, and group members keep granted manage access when the group changes.

• Admin Agent Benchmarking visible to the right people: Agent Benchmarking is shown only to users with the chat.agent-benchmarking.manage or chat.agent-benchmarking.run role, instead of to every admin.

• Admin Accurate telemetry: Message telemetry reports token usage for classic Space answers instead of zeros, including reasoning tokens.

• Admin Cancelled downloads free resources: File downloads cancelled by the client release server resources immediately.

🔒 Security

• Admin LLM proxies restricted to an allowlist: The OpenAI and Anthropic proxies forward only the inference routes Unique uses and reject every other route with 403. Operators can allow more routes (see 2026.42 Infrastructure Changes); API details on 2026.42 Unique API / SDK Changes.

  • Behavior change: calls to other routes such as /v1/assistants, /v1/files or /v1/vector_stores now return 403.

• User Experimental Scheduled tasks check eligibility before every run: A run is blocked when the owner lost access or left the company, or the space or chat is gone. Blocked tasks show the reason.

• Admin Leaner feedback export: The feedback CSV export no longer includes internal message debug data.

• Admin Container images patched: Platform images fix published critical and high vulnerabilities, frontend images no longer ship unused build tooling, and the route block list now also applies to public routes. Details on 2026.42 Infrastructure Changes.

RAG

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

• User Experimental Personal home folder in the Knowledge Base: Each user’s home folder is listed first in the Knowledge Base and is visible only to them. Homes are named after the user, for example /home/JaneDoe, and the old path keeps working. See Knowledge Base for End Users.

• User Experimental Open a file straight to the cited pages: Knowledge Base links can open a file in the preview with the cited pages highlighted, and the Knowledge Base MCP search tool returns this link. See Knowledge Base MCP - Tools.

  • Behavior change: the details pane cannot be dragged wider while a preview is open.

🔧 Improvements

• User GA New Knowledge Base interface on by default: The redesigned Knowledge Base with new list, search and column views is now on by default. See Knowledge Base for End Users.

  • Behavior change: companies that have not set the flag switch to the new interface after upgrading; setting it to false keeps the previous one.

• User Experimental Home folders are protected: Owners can add and edit files in their home folder but can no longer change its sharing, permissions or ingestion settings.

• Admin Experimental Nobody can change home folder sharing: Not even administrators can change the sharing of home folders, and inside a home only its owner can change subfolder sharing. See Knowledge Base for Admins.

• Admin Experimental Extracted metadata searchable from the first ingestion: Metadata from LLM metadata extraction can be added to each chunk during the first ingestion, so no re-ingestion is needed. Requires LLM metadata extraction to be on as well.

• Admin Experimental Protected files ingested with the uploader’s rights: Microsoft Purview protected files are decrypted with the uploader’s rights, so a file is only ingested if that user can open it.

  • Behavior change: operators must change the ingestion worker’s Entra permission and set one variable before upgrading (see 2026.42 Infrastructure Changes).

🐞 Bug Fixes

• User Experimental Chat-only users can get view access: Users who only use chat can be given view access to a Knowledge Base folder and use its content in chat. Applies where folder permission management is enabled.

  • Behavior change: companies with that flag on can now grant view access to chat-only users.

• User Very large table rows ingest: Files with very large table rows, such as CSVs with long multi-line fields, now finish ingesting. Oversized rows are split and keep their column headers.

• User Widely shared files can be updated: Permissions of files shared with about 100 or more users or groups can be updated again, which also unblocks SharePoint site syncs. See SharePoint Connector.

• User Metadata values are kept: File metadata keeps values of 0, false and empty text, also when a new file version is uploaded.

• User Versioned folders handle moves and emptying: Files can be moved into a versioned folder even after a file with the same name was deleted there, and Empty Folder keeps version history.

• Admin PDF ingestion recovers automatically: When PDF processing loses its connection to its job queue, it now restarts instead of silently accepting jobs it never processes.

• Admin Failed ingestions are recorded: Files that permanently fail ingestion are recorded in the ingestion error analytics instead of being discarded. Monitoring changes are on 2026.42 Infrastructure Changes.

• Admin Retention cleanup leaves nothing behind: Expiring files now also lose their generated PDF preview, and chat files that expired under older versions can be fully removed; that cleanup is off by default (see 2026.42 Infrastructure Changes). See Data Retention and Data Deletion at Unique.

🔒 Security

• Admin Container images patched: The ingestor image fixes a published critical dependency vulnerability and two container vulnerabilities. Details on 2026.42 Infrastructure Changes.

Agentic Table

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

• User Experimental Add verified rows to the Answer Library in bulk: Sheet managers and owners can add selected verified rows to the Answer Library in one step.

• User Share a sheet with one person from a group: You can share a sheet with a single member of a group that has access to the space.

• User Conduct agents can fill a single cell: A Conduct agent can write text into one cell without regenerating the row.

• Admin Experimental Turn notifications on or off: Sheet owners and managers can turn assignment notifications on or off for a sheet. Turning them on lists who will be notified, and names the group when someone was assigned through one. See Assignment Notifications (Email, In-App).

• Admin Answer Library metadata labels: Admins can define Answer Library labels, such as Team, in the space configuration without code changes. Off by default.

  • Behavior change: spaces that use the old extract-time tagging keep it until they migrate; that fallback is removed in the next release.

• Admin More control through the Public API: Row deletion can now delete permanently and remove the row’s ingested Knowledge Base content, and bulk cell updates can overwrite existing Answer Library rows. See Unique API Documentation.

🔧 Improvements

• User Experimental Select all search results in a column filter: While searching a column filter, a checkbox selects only the matching values and keeps the rest of the selection.

• User Hierarchical question extraction by default: New sheets extract questions from the full document with hierarchical extraction.

  • Behavior change: spaces that never set the extraction mode now use hierarchical extraction; spaces that explicitly turned it off are unchanged.

• User One row per top-level question: Sub-questions stay inside their question’s row, and attachment and signature requests get their own answer types.

  • Behavior change: documents with nested questions produce fewer rows, and the Section column shows the full path, such as 1 - General > 1.2 - Operations. Admins can restore the previous value with sectionFormat.

• User No duplicate library entries: Adding a row to the Answer Library again updates the earlier entry instead of creating a duplicate.

• Admin Subagent columns use a fresh chat per row: Subagent columns now start a separate chat for each row.

  • Behavior change: columns that do not set subagent_reuse_chat now use a fresh chat per row; set it to true to keep the shared chat.

🐞 Bug Fixes

• User Add to Library ingests each row once: Completing a sheet with Add to Library no longer creates duplicate or re-labelled Answer Library files.

• User Loading indicator during extraction: The sheet shows a loading animation while questions are extracted.

🔒 Security

• Admin Container images patched: The Agentic Table image fixes published high container vulnerabilities. Details on 2026.42 Infrastructure Changes.

Unique AI

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

• User GA Context Memory for every company: Context Memory is general available. Space administrators switch it on per space, and users can opt out in Settings. The feature can be activated/deactivated per company via the feature flag FEATURE_FLAG_ENABLE_CONTEXT_MEMORY_UN_22033. See user and admin documentation.

🔧 Improvements

• User Better visibility of uploaded files for the agent: Each uploaded file now comes with a status line telling the agent whether and how it can be searched, is attached as an image, or is available in the code execution container. This also covers files that are still ingesting, failed ingestion, or expired.

🐞 Bug Fixes

• User Deep Research finish: Completes instead of staying on the progress bar.

• User Code Execution uploads no longer hang: Uploading a file to a Code Execution container no longer hangs for 10 minutes; a stalled upload is retried within about 30 seconds.

🔒 Security

• Admin Container images patched: Assistant images fix published critical and high container vulnerabilities.

Language Model

ADMIN Claude Sonnet 5.5

Claude Sonnet 5.5 is now available via LiteLLM via Anthropic and VertexAI:

• litellm:anthropic-claude-sonnet-5-5

• litellm:vertex-claude-sonnet-5-5

Admin Claude Opus 5.5

Claude Opus 5.5 is now available via Anthropic and Vertex AI:

• litellm:anthropic-claude-opus-5-5

• litellm:vertex-claude-opus-5-5

ADMIN GPT 6.1 Sol

GPT 6.1 Sol is now available via Azure (DataZone Standard deployment):

• AZURE_GPT_61_SOL_2026_0929

Admin Azure OpenAI o-series - Retirement & Migration

Ahead of Microsoft retiring these models on November 19, 2026, spaces, space templates, module configurations and scheduled tasks that use them move automatically:

• Azure OpenAI o1, o3 and o3 Pro move to GPT 5.6 Sol.

• Azure OpenAI o3 Mini and o4 Mini move to GPT 5.6 Terra.

• The OpenAI-hosted o1, o3, o3 Pro, o3 Deep Research and o4 Mini move to the OpenAI-hosted GPT 5.6 Sol and GPT 5.6 Terra the same way.