📌 Release Overview

• 2026.38

• Unique hosted: September 16, 2026

Administration / Governance

🔧 Improvements

• Admin Beta The Cost Management dashboard now provides redesigned usage reporting with period navigation and custom date ranges.

• Admin Reported model spend for GPT-5.4, GPT-5.4 Pro, GPT-5.5, and GPT-5.6 (Sol/Terra/Luna) requests exceeding 272,000 input tokens now reflects OpenAI's long-context pricing: the entire request is billed at 2x input-side and 1.5x output rates.

  • Behavior change: GPT-5.6 Terra and Luna base rates are also updated to the current list prices (Terra 2/12, Luna 0.20/1.20 per 1M tokens).

• Admin Self-registration founding admins receive full Unique roles; new organizations are granted every project role so admins can assign them. Custom password sign-in rejects unverified emails until the mailbox is verified or manually approved in Zitadel. Gated by ALLOW_SELF_REGISTRATION (off by default).

• OPERATOR Self-registration can skip email verification when REQUIRE_SELF_REGISTRATION_EMAIL_VERIFICATION is false (default true). QA sets it false because the cluster cannot send verification mail. Accepted risk: a domain can be claimed without mailbox proof while ALLOW_SELF_REGISTRATION is on.

🐞 Bug Fixes

• Admin the Analytics model usage report now honours the selected assistant filter instead of always exporting usage for every assistant.

• Admin Password sign-in finds the account by email even when the Zitadel login name differs, and new self-registered orgs keep the email domain so later users can join. Requires ALLOW_SELF_REGISTRATION.

• Admin Group member lists and counts no longer include users that were deleted or cannot be resolved.

• Admin Space and chat admins can install platform packages. The grant is applied when Gatekeeper restarts after this deploy; no extra config is required.

• Admin Installing a platform package no longer fails because installer folder paths were missing a leading slash. No extra config is required.

• Admin The Members page now displays service user names correctly.

• Admin Installing a package always shares created spaces and folders with the installer. Additional group access, including All Company, is applied only where the package declares it.

• Admin Admin Skills additional-source chips no longer rewrite the form when the space is read-only or when only the space root id is already in the selection.

API / SDK

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

• User GraphQL mutation to propose a refined Agentic Table cell answer without writing the cell. Disabled by default; enable via FEATURE_FLAG_AGENTIC_TABLE_EDIT_WITH_AI_UN_24386.

🔧 Improvements

• OPERATOR Knowledge base Public API routes (/folder, /content, /chunk, /search, /ingestion-metrics) are now served by the Unique API service. Request and response shapes are unchanged; POST /content/magic-table-sheets continues to be served by node-chat.

🐞 Bug Fixes

• User Public API errors now preserve meaningful upstream HTTP status codes.

• User Content search on API version 2026-03-01 now returns ownerId, matching version 2023-12-06.

• User Structured-output chat completions now preserve parsed responses when using snake_case request options on API version 2023-12-06.

• User Long-running Public API requests now use the gateway timeout, and content pagination supports offsets up to 100,000.

• User The folder list API no longer exposes internal system folders through the legacy API version.

• User LLM proxy requests now block provider-hosted tools that can execute outside Unique-controlled environments.

• User LLM proxy requests that group client-executed tools into a Responses API namespace (as sent by recent Codex versions) are accepted again instead of being rejected as hosted tools.

• User Stock OpenAI-compatible clients (Continue, Kilo, OpenCode, coding agents) that send max_tokens no longer receive a 400 error when targeting a reasoning-model deployment (o-series, gpt-5/6 family) through the OpenAI proxy.

🔒 Security

• Admin: Public API requests now enforce caller roles when accessing downstream services.
  Behavior change: Requests without the required caller role are denied.

Chat Experience

👀 Check status of your chats at a glance

The chat list in the navigation now shows small indicators on chats you own: whether a chat is still working, waiting for your input, or has unread results. This is on for everyone — especially useful together with Scheduled Tasks and longer-running agent conversations.

📄 Better document translation

• Translated PowerPoint and Word documents now keep their original alignment and formatting, use the right typography for the target language, and preserve symbols and artwork that previously could be distorted.

• Admins can configure Translator V2 to reuse approved translations from an existing translation memory (an XLSX file) — so recurring phrases are translated consistently with previously approved wording.

• When the source language of a document can't be detected, the translator now falls back to the glossary's default column instead of skipping glossary terms.

🎨 A refreshed look and feel

• The user menu has been rebuilt with a refreshed header and avatar.

• Notifications (toasts) across chat, admin, knowledge upload, and theming now use a redesigned stacking style — enabled by default.

• When no custom theme is configured, the product now starts from a light color palette, and new themes created in the theme editor start from the same light defaults.

• Side panels (document previews, code output, etc.) now share a consistent header with fullscreen and download actions, and many small inconsistencies in dropdowns, dialogs, and menus were polished: consistent styling, no more menus running off-screen, smooth scrolling in long lists, and branded theme colors applied consistently throughout.

🧩 Skills overview in Settings

• User EXPERIMENTAL Users with the Skills feature enabled get a Skills tab in User Settings listing their skills with search, filters, and pagination — groundwork for broader skills management to come.

🐞 Bug Fixes

• User Agent chats keep their memory: long, tool-heavy conversations no longer lose their session context mid-chat with a false “context window filled up” notice.

• User Answers keep streaming through brief infrastructure interruptions instead of stalling until a restart.

• User Files generated in chat (via code execution) now always display correctly.

• User Deep Research with the OpenAI engine completes runs instead of failing immediately.

• User Interrupted answers are honest: when an answer stops before finishing, users see a clear “stopped early” alert with a Try again action, and aborted steps are shown in red rather than looking completed.

• User Chat sharing: adding a person in the sharing dialog now grants access immediately (previously it only took effect once the link was copied), and anyone on a chat's access list can open any link created for that chat.

• User Document previews: PDF and Word previews in the Knowledge Base load reliably across deployments, and zoomed PDFs stay fully scrollable.

• User The hallucination warning on an answer stays visible while scrolling through a long message.

• ADMIN the Reference Statistics analytics report now respects the selected date range, and the model usage report honours the assistant filter.

🔒 Important change for API consumers

Preview features — reference for enablement requests

Conduct Beta

Highlights of this release: faster first answers in every chat, skills that can be sourced from more than one Knowledge Base folder, stronger grounding and citation quality, more reliable turn completion and stopping, and new usage, cost, and memory reporting for admins and operators.

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

• Admin Claude Sonnet 5 and DeepSeek V4 Pro: Claude Sonnet 5 is a model choice for Conduct spaces. DeepSeek V4 Pro is available for the Conduct Pi Agent (litellm:deepseek-v4-pro).

• User Grok 4.6 on QA: Grok 4.6 is selectable for Conduct spaces and the Conduct Pi Agent on QA (litellm:grok-4-6). Spaces on the Pi harness complete Grok 4.6 requests successfully.

• Admin Personal skills toggle: The Sources card includes a Personal skills switch (enablePersonalSkills, on by default) that controls whether each user's private skills folder is loaded.

• User Fewer false groundedness warnings: The hallucination check no longer flags supported answers. Example-code citations, crawler pages, uncited tool output, multi-entity results, and recommendation questions are handled correctly.

• User The assistant says when it cannot remember: When context memory is off, the agent states that it cannot remember across conversations and names the reason (web search, space setting, or user opt-out).

• User Experimental Metrics tab (only in Conduct spaces): The message debug Metrics tab appears only in Conduct spaces. Other spaces no longer show empty metrics.

• OPERATOR Conduct Usage dashboard: A Grafana dashboard tracks turn volume, harness mix, token usage, feature adoption, and usage-quality signals alongside the SBX Turn Timing dashboard. It shows integer turn and token counts, drops unwired cost and tool-call panels, uses active harness filters from Prometheus, and appears in the apps folder when sbx-templates provisioning is enabled.

• Admin User Memory reporting covers Conduct: Conduct turns report memory load, storage, write outcomes, and latency on shared Prometheus series. Every load and write counts, including the first on a fresh pod, so Loads and Memory headlines no longer show 0 after one chat. The dashboard adds per-range totals bar charts and a Distributions row with histogram buckets for load, write, LLM and storage duration, and profile size.

🔧 Improvements

• User Faster first answer: The agent environment warms on chat open and first keystroke. The first message adopts a pre-started workspace and agent. The agent starts when workspace and user memory are ready, not after history fetch. Warm-up skips the legacy asset prefetch and runs for all harnesses. The extra SDK version check per connection is removed (~0.3–2s saved). Short prompts now get answers several seconds sooner.

• User Connectors start in-process: Connectors and connected spaces run inside the long-lived agent host instead of a fresh process on every start. This removes ~4–6 seconds from the first answer across all harnesses. If the in-process host is unavailable, the agent falls back transparently.

• User Faster skills sync: Skill folders list in batched requests with concurrent layer discovery (~4x fewer platform calls). Pods reuse shared discovery, parallel manifest and folder listing, warm-pool bundled-skills cache, and background revalidation on cache expiry. The agent starts during skills refresh (~1s off time to first token). The assistant launches once skills link from cache. The public content search response includes ownerId.

• User Faster turn end and memory load: User-memory updates no longer hold up the answer (~1–2s saved). Skills refresh reuses cached bundles. Memory loads from the turn manifest and pod-local cache: at most one API call on a cold turn, none on warm turns.

• Admin Skills root required before enable: Skills Settings requires a root folder before Skills can be enabled, shown as an editable chip in the Sources card.

• User Legacy document skills retired: Brand-pack, template-library, and shared document-generation skills no longer appear in the Conduct skill catalog.

• Admin Codex defaults to medium reasoning: New or unconfigured spaces use medium reasoning effort. Existing spaces keep their setting, and the setting remains adjustable per space.

• Admin Agent-sandbox v1beta1 APIs: Sandbox templates, claims, and warm-pool claims use agent-sandbox v1beta1 APIs (controller v1.0.1).

• Admin Clearer User Memory dashboard: Outcome mixes (load, write, gate, scrub, consolidation, storage, condense, errors, LLM calls) show as pie charts. Headline stats show median and 95th-percentile latencies. Latency distributions use labelled buckets and whole-number counts. Panels state what is counted and over which window. Headlines show 0 instead of "No data".

🐞 Bug Fixes

• User Partial answers explain themselves: When a turn stops mid-answer, the interruption warning now appears in the response below the partial text, not only on a hidden step row.

• User Successful turns no longer show errors: Turns no longer end with an unrecoverable error when the result succeeded. Stalled turns report what tool calls or files survived.

• User Interrupted tasks resume where they left off: A turn resumed after mid-task interruption no longer restarts from scratch or re-narrates completed work.

• User Long reasoning turns survive compaction: Turns that reason without visible tool calls no longer abort silently, including during context compaction. Run-stats and Debug/Report Issue still render when a turn aborts with no answer.

• User 1M context window detected correctly: Claude Opus/Sonnet 4.6+ no longer misdetect a 200k window and repeatedly auto-compact.

• User Stalled coordination no longer hangs forever: Turns no longer wait indefinitely for stalled sandbox coordination requests.

• User Tools work in pre-started chats: Web search and other tools no longer fail with "Chat was not found for this user" in pre-started agent chats.

• User Waiting for your answer is visible: The agent shows when it waits for your answer. The wait is no longer cut off by the Bash timeout. Off by default; enable the Claude elicitation_supervisor space setting.

• User Clean answers without duplication: Short or repeated correct answers no longer block the reply or duplicate an earlier turn. Intermediate progress stays separate from the final answer.

• User Each cited source gets its own footnote: MCP connector results no longer collapse onto one reference chip when every result shares the tool name.

• User Unexecuted plans close clearly: A turn that stops on an unexecuted plan lays out the plan and states it was not executed. The generic incomplete notice remains when the plan guard is off.

• User Updated uploads can be downloaded: Rewritten chat-uploaded files download as a chat card under the original filename. The original upload stays unchanged.

• User Generated files always appear as download cards: Files mentioned only in a code transcript or quote appear as a single download card.

• User In-place edits find the right file: The agent knows where attachments landed and what survived in ./output/ from the previous turn.

• User Elicitation checkboxes work on first Confirm: Unchecked boxes submit as unchecked without a check-then-uncheck round trip.

• User Skills invoke reliably: Invoked skills appear in the activity log and run from chat input. skills.md folders appear in the picker and run when invoked; if a skill cannot be found, the assistant says so instead of showing a mismatched "Loaded Skill" notice. Skills with spaces or special characters in folder names work for Claude Code, Codex, and Pi. Selecting two skills that normalise to the same name invokes the first and logs a warning. Codex reads workspace skills before tasks, including GPT-5.1 file generation.

• User Memory and reminders behave correctly: Context-memory loading sits with skill and thinking steps. Update status and review link appear after responses when the profile changes. Scheduled reminders state repeat schedule, stay in the current chat, and confirm only after the task is created.

• User Debug metrics report correctly: Run stats attach to the visible answer. Per-turn metrics (turn count, tool time, model wait, time to first token) are no longer blank. Turn-limit stops report correctly. Codex and Cursor show turns in the same unit as their harness limit.

• Admin Sandbox diagnostics reach logs and dashboards: Claude Code SDK errors log on failure (status code, stop reason, error count). Operational logs (phase markers, counts, turn outcome) reach Loki on the next unique-polyagent image. Interrupt trails log on stalled turns. Time-to-first-token and checkpoint save/restore metrics emit again. Egress gateway replicas scale down after bursts; ships with the next sandbox gateway deploy, no flag.

• User Fewer mid-turn stalls from liveness probes: Probes use a 5-second timeout and six consecutive failures before termination, preventing 10–30-second stalls during active turns.

🔒 Security

• User Nested agent CLIs blocked: Conduct spaces refuse nested Claude, Codex, Pi, or Cursor CLIs from another harness. Always on, with no feature flag or space setting.

• Admin Hardened sandbox storage: Checkpoints and skill caches are protected against crafted paths and symbolic-link traversal.

• Admin Patched container images: Conduct sandbox images patch npm and Cursor SDK dependencies for published CRITICAL and HIGH vulnerabilities, plus further published container vulnerabilities. The Assistants Core image patches two published critical vulnerabilities.

MCP Hub Beta

🔧 Improvements

• User Chat: My Connectors now lists only the connectors you have connected; connecting is done from the chat tools menu.

🐞 Bug Fixes

• User My Connectors now shows connectors as connected when the access token has expired but a refresh token is still valid.

• User Failed MCP server connection attempts no longer replace the chat with an error page. Users can now retry from the interface.

• User The Connect or Reconnect banner now appears when an MCP connector has no or expired OAuth token, instead of a generic processing error.

• Admin Virtual server endpoint slugs are unique across companies. Creating a server whose slug is already taken is rejected.
  Behavior change: Existing duplicate slugs keep the oldest value; later servers are renamed with a numeric suffix (-2, -3) and their public MCP URL changes. Creating a virtual server with a slug that already exists now shows a clear conflict message.

Agentic Table

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

• User Experimental Edit Agentic Table answers with AI: Users can refine an answer cell in an ephemeral chat and accept the result. Disabled by default; admins enable it per company via FEATURE_FLAG_AGENTIC_TABLE_EDIT_WITH_AI_UN_24386.

• User Experimental Stop an in-progress table run: Users who can edit an Agentic Table can Stop an in-progress run. Confirming halt keeps cells already written, starts no further cells, lets in-flight work settle, skips unextracted questions, and returns the sheet to idle with a stopped-by-user note. Disabled by default; enable via FEATURE_FLAG_AGENTIC_TABLE_STOP_REFRESH_UN_11560.

• User Beta Bulk actions are now beta: After UX improvements, Agentic Table bulk actions are in beta. Disabled by default; enable via FEATURE_FLAG_AGENTIC_TABLE_BULK_ACTIONS_UN_23981.

• User Beta Bulk add or remove row metadata: Bulk actions on Agentic Table now include an action to add or remove row metadata on the selected rows. Verified rows are excluded. Disabled by default; enable via FEATURE_FLAG_AGENTIC_TABLE_ROW_METADATA_UN_14904 together with FEATURE_FLAG_AGENTIC_TABLE_BULK_ACTIONS_UN_23981. The same update is available via the GraphQL mutation bulkUpdateMagicTableSheetRowMetadata (existing keys overwritten; removeExistingMetadata replaces or clears all tags; up to 50,000 rows).

🔧 Improvements

• User Outdated configuration banner: Invalid or outdated Agentic Table configurations now appear as a dismissible banner, with a link to edit the space configuration for admins.

• User Bulk assign can now unassign: Bulk actions use the updated dialog design. Bulk assign can remove all assignees from selected rows when Remove existing assignees is checked with no new owners; the modal action is labeled Confirm for both assigning and unassigning.

• User Change assignees on Verified rows: Assignees on Verified rows can be changed without unlocking question or answer content.

• User Loading skeleton on re-run: When rows are re-run, each affected answer cell shows a loading skeleton until the new answer arrives. Question cells stay as they are.

• User Larger bulk review-status updates: Bulk review-status updates now accept selections of up to 50,000 rows; rows that already have the chosen status are left unchanged. When more than 1,000 rows change, the table refreshes once the update finishes rather than updating each cell live.

• User Server-side Full Report Excel export: Full Report Excel export now runs on the server, with progress for large tables. Markdown formatting in Full Report XLSX is on by default; set xlsx_generator_config.render_markdown_on_export to false in the space config to export plain text.

• User Sheet metadata form on table creation: Creating an Agentic Table now shows the sheet metadata form when it is configured in the space’s advanced settings, except for Source of Wealth and Engagement Digest onboarding. The FEATURE_FLAG_AGENTIC_TABLE_METADATA_FORM_UN_17903 gate is removed, so companies that previously had it off now see this form if it has been configured.

• User Table question grids and choice options: Table questions now keep the questionnaire’s grid in question details, and preserve per-column dropdown, radio, and checkbox choices there and in Word exports. The grid is on by default and replaces the previous “Rows:” and “Columns:” lists; set includeTableMarkdown to false on a space to keep those lists. This needs additionally the following config settings to be set to true: useFullDocument , useHierarchicalExtraction, and useXlsxPreprocessing.

🐞 Bug Fixes

• User Tables & references while editing a cell: While editing an Agentic Table cell, tables and source references stay visible, and pasting a markdown table together with surrounding text now renders the table instead of leaving raw markdown.

• Admin Answer-library tagging when turned off: Turning off answer-library tagging no longer writes client, strategy, or as-of tags onto extracted rows.

• User Excel formulas, dates, and percentages: Excel questionnaires no longer treat formulas as content, and dates and percentages now appear as the spreadsheet shows them rather than as Excel’s stored values. A total such as =SUM(C6:C10) is left out of the question; a column headed “Mar-26” reads as “Mar-26”, and “-0.74%” stays “-0.74%”.

• User Live review status and row locks: Review status and row lock changes now appear immediately for other people viewing the same sheet.

RAG

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

• Admin EXPERIMENTAL Data administrators can reveal hidden system Knowledge Base folders, such as the per-user memory home folders, with a Show hidden folders toggle in the Knowledge Base admin. Off by default. Enable via the flag FEATURE_FLAG_ENABLE_SHOW_HIDDEN_FOLDERS_UN_24840.
  

  

• Admin Microsoft Purview (MIP) protected Office files can be decrypted during Knowledge Base ingestion. Off by default. Enable it with ingestionConfig.mipDecryption.enabled on a company, scope, or space, and set MIP_DECRYPTION_ENABLED plus MIP_AZURE_TENANT_ID, MIP_AZURE_CLIENT_ID, and MIP_AZURE_CLIENT_SECRET on the ingestion worker.

• User The internal file picker now shows the total number of selected files.

• Admin Custom file metadata can be embedded together with each chunk so it influences vector search similarity, not only filters. System-managed fields (key, title, folder and content ids, mime type) are never embedded. Off by default. Enable it with ingestionConfig.chunkMetadata.enabled on a folder, space, or company (checkbox 'Add file metadata to chunks' in the ingestion configuration, or via the folder ingestion-config API); ingestionConfig.chunkMetadata.maxTokens (default 256) caps the size of the metadata block. Files ingested before enabling it must be re-ingested to pick it up; metadata produced by LLM metadata extraction is only included after a re-ingest.

🔧 Improvements

• User Qdrant searches use the universal Query API, and new collections can use Qdrant 1.19 memory tiers. Memory tiers are disabled by default; enable them with QDRANT_USE_MEMORY_CONFIG only after every Qdrant node is on 1.19 or later.

🐞 Bug Fixes

• OPERATOR The ingestor's Hugging Face CDN egress allow-list now matches any region's model-download host instead of only specific subdomain depths.

• User Metadata-filtered content queries no longer return soft-deleted files, and moving files now updates the searchable folder metadata immediately instead of only after the deferred rebuild.

• User File-based search now applies access inherited from parent groups when FEATURE_FLAG_ENABLE_FILE_BASED_ACCESS_UN_12660 is enabled.

• OPERATOR Recency-ranked searches on large knowledge bases now complete without timing out when RECENCY_RANKING_COMPANY_IDS or recencyBoost.enabled is configured. New Qdrant collections receive a datetime payload index on metadata.validAsOf, and a data migration backfills the index for existing collections.

• User The Knowledge Base file details Metadata section now shows numeric and boolean metadata fields (for example IDs, sizes or yes/no flags produced by metadata extraction) instead of silently omitting them, and the item count reflects every field shown.

• User Restricted Knowledge Base ancestors now show the correct subfolder count.

Language Model

ADMIN Grok 4.6

Grok 4.6 is now available through LiteLLM via xAI:

• litellm:grok-4-6

ADMIN GPT 6 Astra

GPT 6 Astra is now available through Azure (global deployment only):

• AZURE_GPT_6_ASTRA_2026_0903

ADMIN o3-mini - Retirement & Migration

Due to the upcoming retirement of OpenAI o3-mini on October 1st 2026, all spaces and AI templates previously using AZURE_o3_MINI_2025_0131 have been migrated to o4-mini (AZURE_o4_MINI_2025_0416).

ADMIN GPT-4o (2024-0513) - Retirement & Migration

Due to the upcoming retirement of OpenAI GPT-4o (2024-0513), all spaces and AI templates previously using AZURE_GPT_4o_2024_0513 have been migrated to GPT-4o (2024-1120) (AZURE_GPT_4o_2024_1120).