What we shipped recently

Improved chat experience

The chat copes better with connection drops, shows more clearly what the assistant is doing, and is quicker to type into:

• Answers continue after a dropped connection: If the connection drops while an answer is being written, the chat shows that it is reconnecting and continues the answer where it left off. If an answer can't be recovered, the chat says so directly below it and offers to retry or switch to another model.
• Clearer tool steps: While the assistant works, the header above its steps shows what it is doing right now. Each step is a single compact row with its search query or target, and its details open in place.
• Each result shown once: Documents, presentations and files created in a chat appear as a single card after the tool steps, instead of one entry per intermediate step.
• Click anywhere to type: The chat input and workflow step prompts take focus from a click anywhere in their box, including the empty space next to the model selector. After New Chat or opening a chat, the cursor is already in the input.

Guided onboarding for new members

New members are guided through a short onboarding before their first chat: they choose their language, enter their name and role, and see which of the company's systems MeinGPT can connect to. Administrators of a workspace that is not set up yet also add its details there. Members who have already completed the welcome tour keep working as before.

More improvements

• Easier navigation: The project page, project data, use case editor, image history and the Learn pages (guides, posts, learning paths, courses and workshop management) show a trail at the top that links back to the parent page. Project settings open from the More actions menu next to the project title.
• Labels on icon buttons: Icon buttons across the app show a short label in your language on hover.
• Compact dropdown menus: Dropdown menus are more compact and open with a quicker, smoother animation.

Bug Fixes

• Deleting a team also removes its link to the Entra group, together with the memberships the sync added to it.
• An assistant picked with @ in the chat input answers with its own instructions and knowledge.
• When a message can't be sent, the chat names the reason: missing permission, no active workspace subscription or no assigned seat. An expired session leads to the login page.
• The toolbar under the chat input moves extra options to a second line when many are active, so the model name and all buttons stay readable.
• If a meeting summary can't be created, the meeting still finishes with its transcript, and the summary can be generated again from the meeting.
• A live recording stays intact when Start is clicked more than once, and the button shows while the recording is starting.
• SharePoint knowledge sources sync every file the connected account can read, even when the site also contains libraries or files that account can't open, such as files awaiting approval.

Gemini 4 Argon

Gemini 4 Argon, Google's most capable model, is now available. It is hosted in the EU and handles very long inputs with a context window of 1 million tokens. Administrators can enable it under Settings → Models. Until November 30, 2026 it costs €2.64 per 1M input tokens and €13.20 per 1M output tokens; from December 1, 2026 the price is €5.28 and €26.40. Details are on the pricing page.

Model retirements

As announced, some models are switched off from October 18 and replaced by newer ones. Assistants, workflows, chats and default model settings that used them were moved automatically to the recommended replacement:

• o4-mini → GPT-5 mini
• o3-pro and GPT-5.4 Pro → GPT-5.6 Sol
• GPT-5.1 and GPT-5.3 Chat → GPT-5.6 Terra
• Claude Haiku 4.5 → Claude Haiku 5.5
• Gemini 2.5 Flash → Gemini 3.5 Flash-Lite
• Gemini 2.5 Pro → Gemini 3.5 Flash
• Gemini 3 Flash and Gemini 3.1 Pro → Gemini 3.8 Flash
• Kimi K2.5 → Kimi K2.6

Wherever a retired model was enabled, its replacement is now enabled too, so existing assistants and workflows keep working. MeinGPT Web Search now runs on Claude Haiku 5.5 and costs considerably less.

Chat improvements

Answers in chat behave more predictably when something goes wrong:

• Stopping keeps what was written: After stopping an answer mid-generation, the assistant remembers the partial answer, so follow-ups like "make that shorter" refer to what was on screen.
• Interrupted answers are recognized sooner: When an answer stops arriving, the chat shows within a few minutes that it was interrupted and offers a retry. The chat shows right away which tool the assistant is using.
• Clearer errors: A failed answer shows a single, clear error message. A timeout during a long answer says so and suggests trying again. Buddy shows the reason directly in the conversation, with a Retry button where trying again can help.

Automatic data source search is now a choice

Chats without an assistant now search all data sources only when this is switched on, and it starts switched off in every workspace. This keeps answers focused on the question and saves credits on searches that would not help. Administrators decide under Settings → Data sources whether automatic search is on for everyone, off, or up to each member; members switch it on in the chat's + menu or under Preferences. A data source picked in the chat or attached to an assistant is always searched.

Deletion periods

Under Settings → Deletion periods, chats, meeting recordings and automation runs can be deleted automatically after a set time. This is off by default. Before a new or shorter period takes effect, there is a 24-hour waiting time, and the setting shows how many items will be deleted.

Workspace administration

• Partner programme: Workspaces can invite their implementation partner with a partner code and accept or decline partner requests. A workspace prepared by a partner is handed over on accepting the invitation, with an overview of what the partner can and cannot see. Administrators see which partner staff have access and until when, decide what they may do, and can revoke access or end the partnership at any time. An email goes out whenever a partner is added. Partner staff do not use up any of the workspace's licences. Applying to become a MeinGPT partner yourself is possible under Settings → Partner.
• Usage exports in the background: Usage exports keep running after refreshing or leaving the page. Progress is visible, an export can be cancelled, and finished files stay available for 24 hours. With anonymized statistics enabled, the export shows a pseudonym per user instead of names and email addresses and leaves out comments on credit bookings.
• Mandatory trainings along learning paths: Mandatory trainings follow the structure learners see in the Academy: a whole learning path, single steps, or courses outside any path. A department-specific course counts as done once any version of it is completed, so members only take the version for their own department.
• Wider tables: The audit log, on-premise search log, governance and members pages use the full width of the screen, so wide tables show more columns at once. In a narrow window the table scrolls sideways with a visible scrollbar.

More improvements

• Clearer navigation: The chat header shows where a chat belongs, "Project / Chat" or "Assistant / Chat", with a link back, and always shows the chat's name. Research, Image Generator, Buddy, scheduled tasks, project workflows and the assistant import page use the same header with a link back to where you came from.
• Connectors stay on: A new chat starts with the connectors used last, so they no longer have to be switched on every time. A bar below the chat input suggests connecting Microsoft and Google apps such as Outlook, Teams, Gmail and Drive, and can be hidden.
• Polish translations for assistants: Assistants can now have a Polish name, description and conversation starters for users who chose Polish.
• Better example assistants: The example assistants come with clear instructions and the right tools already switched on. A new Assistant Creator example guides step by step through building your own assistant.
• Meeting limit shown in advance: When the meeting limit is used up, the Meetings page says so before the next meeting, shows when it resets and offers to request a higher usage tier.
• More readable colours: Error and delete states are darker, informational and in-progress states use one consistent blue that stays readable in dark mode, and selected or hovered items use a neutral or brand-coloured highlight.

Fixes

• Search in large files attached to assistants and chats now takes the whole file into account.
• Searching Microsoft Teams channel messages is more reliable, in one channel or across all channels of a team, including private channels.
• The customs tariff (TARIC) search reliably finds tariff codes and customs offices, and a commercial register search without matches shows an empty result.
• The privacy filter also recognizes account numbers that wrap across a line, puts the original names back correctly in summaries of scheduled tasks, applies the chat's filter setting to chat titles, and handles long requests more reliably.
• Web search sources show umlauts and special characters correctly in their titles.
• Meeting settings and upcoming meetings load reliably when the notetaker or auto-join is switched off.
• Large member lists and member selection fields load profile pictures faster.

For administrators

• Single sign-on per workspace: Workspaces can offer sign-in through their own identity provider (OpenID Connect, e.g. Keycloak or Okta). The login page asks for the email address first and then shows the sign-in methods available for that workspace. To enable this for a workspace, contact the sales team.

Fixes

• Stability and response times under high load are improved.
• Images returned by tools during assistant runs via the API, automations or workflows reach most models as images instead of text, so these runs use far fewer credits. In the API response, such tool results list the image as a content part.

Claude Haiku 5.5

Claude Haiku 5.5, Anthropic's new compact model, is available in chat and assistants. It is hosted in the EU and supports adjustable reasoning. Administrators can enable it in the model settings.

Model retirements

Following Google's retirement, Gemini 3.6 Flash will be switched off on November 13, 2026 and Gemini 3.7 Flash on January 22, 2027. Assistants and workflows that use them move to Gemini 3.8 Flash.

For administrators

• A page for every member: Clicking a name in the member list opens a page with role, usage group, teams and an access check that shows what the member can use and why. On the shared content page, older projects can now be reviewed and handed over too.
• Connecting follows "Usable by": Members can only connect a connector they are allowed to use under "Usable by"; using a restricted connector was already blocked. Members with read-only access are asked for read permissions only, where the connector supports it.

More improvements

• Usage by model: Personal usage limits show a separate progress bar for each category and, if an administrator enables it, a breakdown of spending by model.
• Meeting bots are more punctual: Bots now join a scheduled meeting about a minute before it starts, and the meeting card shows when a bot is on its way. After the last participant has left, the bot waits three minutes before leaving.

Fixes

• Images and files uploaded to a chat reach the model that answers, even when the chat switches to a different model. If that model cannot view images, the answer says so.
• Outlook drafts and emails written by an assistant can include files uploaded to the chat, and the attachments open correctly.
• Searching Outlook for unread or flagged emails on a specific topic works and no longer adds unwanted filters.
• Outlook search shows attachments, and they can be added to a data pool, including from emails with cloud-file links or attached emails and invites.
• The meeting bot can be switched on by hand for Outlook meetings whose invitation came from another calendar, such as Google Calendar.
• Knowledge search covers large workbooks with many sheets more completely and finds calculated prices and amounts as they appear in the sheet.
• Knowledge search matches amounts written with either a comma or a dot as the decimal separator.
• When an external connector such as Make, GitHub, Asana or HubSpot is disconnected or reconnected, MeinGPT now also withdraws the old authorization at the provider, so it no longer counts against that provider's connection limit.
• Deleting an unfinished assistant when leaving it no longer keeps the deleted draft open

Fixes

• Every sheet of a spreadsheet is searchable — Knowledge search finds rows in all sheets of a workbook, not only in the first ones, and finds calculated prices and amounts as they appear in the sheet. This applies to spreadsheets added or updated from now on.
• Amounts with comma or dot — Knowledge search finds an amount whether it is written with a comma or a dot, for example 1.234,50 or 1,234.50.

Fixes

• Exact numbers and codes are found — Knowledge search finds passages that contain a specific order, ticket or part number, even when the question is a full sentence. Passages also rank higher when the question and the document are in different languages.
• Fewer failed files — Files that hit a temporary outage while being indexed are retried for several hours instead of being marked as failed.
• Outpost folders show their real state — A shared folder that was moved, deleted or whose drive was disconnected now shows as unavailable, and "last sync" means the last complete scan. Deleting a folder also removes its content from keyword search.
• Replaced files are read again — A file that is replaced on an Outpost while it waits to be indexed is read again, so search results always match the current file.
• Clear errors from the embeddings API — When the API rejects an input, for example because it is too long, it now answers with a 4xx error that names the problem instead of a server error.

Faster search

Knowledge search runs on a new index and is up to 15× faster, so answers from knowledge sources arrive noticeably sooner.

SharePoint sync works again

SharePoint sources connected through a Teams site address or a link into a document library sync again. Since the previous update they had stopped picking up changes; no content was lost.

Scheduled tasks stay with their person

Scheduled tasks run, and show their results, only for the person they run as. Colleagues who edit the assistant still see a task's status and can switch it off or delete it.

Knowledge search in every workspace

Chat searches all available sources at once and shows each match with its file, page and section. Cmd+K finds files by name and can be filtered by source, file type and date.

Pictures in PDFs become searchable

Search now also finds pictures and drawings in PDFs, and answers show them directly.

Owners for data sources

Administrators can name people or teams who manage a data source without admin rights.

A more stable Outpost

Supports Windows Server 2016, works behind a proxy and starts with 200 GiB of storage.

More improvements

• SharePoint, OneDrive and Google Drive sync more reliably.
• Every Cloud source uses meaning-based search.
• The AI Cockpit shows how data sources are used.
• Generated images are marked as AI-generated.
• A MeinGPT account stays tied to the Google or Microsoft account it was first connected to.
• Numerous bug fixes.

New models: Claude Sonnet 5.5 and GPT-6.1 Sol

Claude Sonnet 5.5 and GPT-6.1 Sol are now available. Both are hosted in the EU and handle a context window of about 1M tokens, so long documents and extended conversations fit in a single chat. Administrators can switch them on in the model settings.

Knowledge search in chat

Chat can now search all connected data sources without selecting one first. Ask a question and the answer draws on every source available to you, ranks results across Cloud and Outpost, and shows which documents it found with links to the originals. Identical passages from multiple files appear once. Rolling out per workspace.

Documents in the search palette

The search palette (Cmd+K) finds files from connected data sources by name and opens the original. When nothing matches, a single Enter sends the question to a new chat that searches your data sources. Results and sources show whether a document is the current or an older version and when its source last synced; clicking a source opens the cited passage. Rolling out per workspace.

For administrators

• Honest progress and alerts — Data sources show the current processing step, searchable files, a rough time estimate, the queue position, and when a sync is stuck. Unreadable files explain why, and administrators get an email and in-app notification as soon as a source needs attention.
• Access review and permissions — Administrators see which data sources are shared with no one, and can share each with a team or everyone, or mark its sharing as reviewed. Source settings state plainly who can find which files; deleted files stop appearing in answers immediately, and access is recorded for security review.
• Improved connect flow — Adding a data source tests the connection first, names what went wrong and shows how many files it found. Sources are named at creation, duplicate folders are flagged, and reconnecting keeps existing settings and starts syncing right away.

More improvements

• Better German search — Knowledge search now finds German word forms, spellings with or without umlauts, and spreadsheet row contents such as part numbers. New data sources benefit right away; existing sources follow gradually as they are re-indexed.
• Gmail delete moves to Trash — The Gmail tool's delete action now moves emails to Trash, where they can be restored. The action is off by default and asks for the matching Gmail permission only when switched on.
• Scheduled tasks pick a valid model — New scheduled tasks start on a model that is actually enabled, and tasks whose model was switched off can be saved again.

Fixed

• Knowledge search prefers the current version of a document over archived or older copies. Answers name data sources by their display name, and files that cannot be previewed or downloaded explain why. On mobile, the sidebar no longer covers the chat.
• Outpost starts even when its search index exceeds the memory budget; keyword search remains available, and the console explains when semantic search is unavailable.
• Files that repeatedly fail during processing show a clear failure instead of staying pending indefinitely. Temporary file-access problems continue to retry, and updates to document processing give failed files another attempt.
• Semantic search in large Outpost folders uses smaller index sections that fit within the memory budget. Search results from smaller folders remain available alongside large collections.
• Outpost continues processing documents during sustained search traffic. Brief CPU spikes no longer pause processing; sustained high load still pauses it automatically.

Meeting summaries without taking notes

The MeinGPT notetaker joins Google Meet, Microsoft Teams and Zoom meetings, records them, and turns the recording into a transcript and a structured summary: topics, decisions, next steps and open questions. To send it into a meeting, paste the meeting link under Meetings → New meeting → Invite bot, add bot@meingpt.com to the calendar invite, or connect a Google or Microsoft 365 calendar so it joins scheduled meetings automatically.

Summary by email to everyone invited. After a recorded meeting, the organizer and the invitees can automatically receive the summary by email, so nobody has to write and send minutes. Administrators choose who receives it, and everyone can narrow this further for their own meetings.

The notetaker is available to every workspace that has signed version 2.0 of the MeinGPT data processing agreement. Administrators sign it under Settings → Contracts & DPA. Workspaces already using the notetaker keep it as is.

Models per team and person

A model that only one department needed used to be switched on for the whole company, or not at all. Now each model can be enabled or disabled per team and person: for example, a specialized coding model only for the development team, or everyone except the trainees. Administrators set this on the AI Models page or directly on a team's page, and each member's page shows which models that person can use. Image models work the same way.

Additionally, Apps, features and connectors can also be limited to the teams and people who need them, including who may let a connector write, for example send emails or create tickets. On top of that, workspaces can decide who may create and share assistants, projects, workflows, automations and API keys, or publish them for the whole company. This is switched on per workspace by the MeinGPT support team. When a member is missing something, their page explains why. (This feature is currently in beta and can be enabled by the support upon request)

More improvements

• Data source status at a glance — The data sources page shows which sources need attention and what to do next. Each source's page explains why a sync failed or a file isn't searchable, shows indexing progress, and offers a test search to see what a colleague would find.
• Settings are easier to find — Settings are grouped by topic. Models, apps and connectors sit side by side, the personal data filter has its own page, and API logs are part of the audit log.
• Simpler data source settings — Options that did not affect search results are removed.
• Outpost search in protected folders — On Windows, Outpost live search returns the files a person may read, even when the folder also contains files they cannot open.

Fixed

• Deleting a knowledge source no longer gets stuck, and crawl filters for websites are applied as configured.
• Answers list only the sources they actually cite.
• Syncing a knowledge source stays possible while it is still being indexed.

Fixes

• PDF citations open the cited page — Citations with page information open the PDF directly on the cited page. Original files download through the browser with their original filenames. Documents indexed before this release show the page once they are processed again.