What we shipped recently

What's new

Admins can set how many tool steps a regular chat may take

Long tool-heavy chats could run longer than a workspace wanted — or stop too early for complex work. Workspace admins can now set the tool step limit for regular chats to 5, 20, or 50 in Chat Settings. Assistant-specific limits stay as they are.

Opening an assistant

Starting a chat from an assistant used to create an empty entry in chat history right away — even when nothing was written. The empty screen still opens, but the chat only appears in history after the first message is sent.

Data Sources

File and chunk counts in data pools no longer drift upwards over time. Downloading original files works for large files and shows the real cause when a connected source system is unavailable.

Large PDFs keep their headings and tables during import. When a file hits a processing limit, that file stops on its own so other documents can continue.

For administrators

• The data source overview loads indexing progress faster and no longer slows down as the number of data sources in the workspace grows.

What's new

Buddy suggests what to do next

Working out what Buddy could help with on a page used to mean typing it out from scratch. Buddy now looks at what's on the page — assistants, projects, workflows, and recent chats — and offers a short list of things it could do, in a small bubble next to its face. One click starts a Buddy chat already set up with that suggestion.

Suggestions can be turned off or refreshed at any time from account settings under Buddy.

Data Sources

Documents that reuse the same pictures — letterheads, template graphics, recurring diagrams — are indexed faster, because a picture a data pool has already read is no longer sent to the vision model again.

Learn

The Responsible AI path adds two lessons — recognising bias (from the AI and yourself) and spotting prompt injection — plus sections on copyright, anonymising data, and reporting incidents, in all six languages. Reflection questions are now optional and stay private.

For administrators

• Low-credit warning emails resume correctly after a top-up: invoice-billed workspaces are warned again the next time the balance runs low, instead of the alert firing only once.
• Workspace settings now keep the configured default model when other settings are saved, instead of quietly resetting it. Chats no longer fall back to a model the workspace has switched off.

What's new

A tidier row for chat attachments

Attachments used to wrap into a grid of tall tiles that ate the space above the message box. They now sit in one scrolling row of shorter tiles. Dropping a file that is already attached says so and points at it. Hovering an attachment shows whether a click opens a preview or downloads the file, and images open in place.

Data Vault

• Scanned documents whose pages are stored as transparent images are now read correctly and become searchable; previously their image content could be indexed as empty.
• Answers from crawled websites are grounded in the site's real pages only. Error pages and repeated copies of a site's start page are no longer indexed as if they were content.

For administrators

Governance

Workspace admins can be given a new Governance page: review every shared assistant, workflow and project, hand ownership over when a colleague leaves, and adjust who has access. Verified ones now carry a seal everyone can see.

Partner sponsorship

Partner admins can create up to 500 open sponsorship links in one step, copy or download them as CSV, and browse large batches grouped by name.

Outpost

• Network folders: downloads recover from slow or unavailable shares, writes stay controlled without duplicate retries, setup checks the real service account before it completes, and large folders stay current without rereading every unchanged file.
• Update channels switch without reinstalling, and production releases use the exact signed Windows, Linux and macOS builds already verified in beta.
• Windows: a fresh install starts as a background service, the administration console opens while that service is running, local data permissions are preserved (and repaired on update if needed), and upgrades continue when an older uninstaller has already removed the app but still reports a failure.
• macOS updates install, close and restart on their own once the verified update is ready, and recover automatically if the new version cannot start.

What's new

A cleaner look for the chat

Headings in a chat reply used to look as loud as the page around them, code sat on the same background as everything else, and sub-points under a numbered step didn't line up with it. Chat now renders with calmer typography: headings step down instead of dominating the answer, code blocks sit on a lighter surface of their own, and sub-points line up under their step. Replying to a passage now shows it as a quote above the message instead of raw markdown text inside it.

Data sources

• Pages behind a login are no longer at risk of being dropped from answers because of how the website spells its sign-in address.
• Cloud Vault status and search checks stay responsive as the number of connected data sources grows.

Bug Fixes

• Chats in the sidebar stay ordered by your most recent message — background activity, like cleaning up an interrupted response or changing a chat's settings, no longer bumps a chat to the top or shows a misleading last-updated time.
• Starting a new chat from a project now includes the Thinking control, so the first message uses the chosen thinking level just like chats started from the home screen.

What's new

Academy courses in Simplified Chinese

The foundational Academy courses — from "what is AI" through writing effective AI instructions to role-specific exercises for every department — are now available in Simplified Chinese. Learners who set their language to Chinese can work through all 42 of these courses in their own language.

Fixed

• German-language searches now find photos, charts and drawings without readable text — image descriptions follow the document's language instead of defaulting to English.
• Sending or regenerating a chat message no longer re-saves every visible email draft, which used to slow down chats with several drafts open at once.
• Course titles in the Academy no longer carry leading numbering like "4.2" or "0." — the names now read cleanly.
• Chats with large PDF or image attachments continue through multi-step tool runs without repeated attachment-processing delays.
• Long or fast-streaming AI responses are less likely to make the chat freeze or crash mid-reply.

Fixed

Website and folder data sources

• Answers drawn from crawled websites and intranets can again include the links and download URLs found on the source page, and are split along the page's own headings for more precise retrieval.
• Answers about crawled websites can again point to files listed in a page's download section; those links were previously missing from the indexed text.
• Pages from crawled websites and intranets no longer vanish from answers when a crawl happens to miss them. A page is only removed after the website itself repeatedly confirms it is gone, and unexpectedly large removals are held back for review.
• Password-protected website sources that need JavaScript rendering now stay signed in for the whole crawl, so member-area pages are stored as real content instead of the login screen.

Connectors

• Google Drive connectors restricted to a folder with many subfolders now list and search files correctly instead of failing with an error.
• The Atlassian connector (Jira, Confluence) no longer disconnects after a few hours; connections now refresh themselves in the background.

Buddy

• Buddy now recognizes every page of the platform and navigates straight to the right one, instead of occasionally landing on a broken link.

Account and support

• Sponsored invitations now activate even if your trial has already expired, instead of stopping with a message to contact support. Invitation links that are no longer valid are explained instead of silently ignored.
• Support notification emails are delivered more reliably when an email provider request stalls or its acknowledgement is lost, without sending duplicate messages.
• The audit log now uses the full width of the screen, so every column — including each change's outcome — is visible without scrolling. In a narrow window the table scrolls sideways and keeps a visible scrollbar.

What's new

• Ask about just what you selected: select any part of an answer and ask about that passage. The highlighted text appears as a quote above the message box, so the follow-up stays focused on what you picked instead of the whole reply.
• Writing clear prompts: the "prompting" foundations course is now "Writing clear prompts" — brief the AI like a new colleague, with a before/after gallery showing how to fix a weak result with one targeted message.

Fixed

• Website data sources no longer show a stuck pending file that was never a document, and large sites keep resuming from where the crawl left off.
• Text in sideways or upside-down document pictures is now read the right way up, so tables and addresses in those images are indexed instead of dropped.
• Website sources recover on their own after a bad crawl instead of keeping the wrong page, and no longer fill up with copies of a login page.
• Password-protected website sources now stay signed in for the whole crawl, so pages are no longer stored as the login screen.
• Sources finish indexing much faster — on a large crawled website, indexing went from roughly 40 to several hundred pages per minute.
• Fixed a bug where re-indexing a large website could remove pages that were still online; the pages are restored on the next crawl.
• Images in a data pool are easier to find by a place or role named in the file, and a missed description is retried instead of staying blank.
• Downloading an original file from a connected storage pool no longer fails when that file isn't already cached.
• Listing and searching Google Drive files no longer fails with a permission error on some AI models.
• Kimi models no longer invent values for optional tool parameters, which used to cause tool calls to fail.
• Custom MCP connectors that loaded none of their tools — such as OpenProject — now work correctly.

For administrators

• Model pricing tiers: premium and cost tiers are now calculated consistently from provider cost, so which models count as premium (billed on top of your plan) may have changed for a small number of models.
• Plan upgrade dialog: the dialog under Settings > Subscription now fits any window height, so the Business and Enterprise buttons stay visible instead of sitting off-screen on smaller displays.
• Data source sync status: the last-sync time on a cloud data source now shows when indexing last succeeded, so a later failed attempt no longer looks like a fresh sync.
• Translator toggle: the Translator can be switched back on after being turned off — its row now stays in the AI apps settings in an off state instead of disappearing.

Say hello to Buddy

Getting help in meinGPT used to mean leaving the page you were on to ask in chat, then carrying out the steps yourself. Buddy is a helper panel that sees the page you're on and helps right there.

Buddy works in three modes. Auto lets it decide how to help, including answering from the documentation without taking any action. Learn shows you the steps with a visible cursor, clicking and explaining as it goes. Act carries out the steps for you — setting up teams, budgets, and models, configuring apps and knowledge pools, or working through learning content.

Every action Buddy takes lands in a history you can review, and anything it does can be undone with one click. If you have any questions, please reach out to support at support@meingpt.com.

Other improvements

• Enter now sends only when you want it to: a new preference lets Enter add a line break instead of sending, so longer messages go out only when you press the send button.
• EU-hosted Gemini 3.6 Flash: prompts sent to this model are now processed only inside the EU.

Fixed

• SQL database tools now open the current custom database documentation instead of a removed page.
• Sponsored invitations now activate on an existing trial workspace instead of stopping at login, so sponsored access arrives without creating another account.
• Running out of your monthly text allowance now shows a short notice naming the model you were switched to, instead of changing the model silently.

What's new

• Meetings, refreshed: starting a new meeting now offers audio, transcript, and paste as tabs, and upcoming meetings show as a week list. Deleting a recording names it and says what is lost, and pasted transcripts no longer claim to have no transcript.

More reliable reading of scanned documents

Scanned PDFs and image-heavy files — technical drawings, tax forms, rotated scans — used to be a weak spot: pages were skipped, garbled, or dropped during text recognition.

Data pools with technical drawings or scanned images can now be indexed with a vision model instead of standard OCR, reading part numbers, dimension tables and tolerances correctly. Rotated pages are read in their displayed orientation, dense pages get the time they need, and documents that were previously rejected are indexed instead of skipped. Re-reading a single scanned document now works reliably too.

What's new

• OAuth for MCP servers without auto-registration — Enter the client ID and secret from the server's admin console once; endpoints fill in automatically where the server publishes them, and each colleague signs in with their own account.
• Search now names what it couldn't apply — If a knowledge source can't honor part of a search — a folder filter, a required term — the assistant says so instead of answering from a wider result set silently.
• New Academy path: Finding use cases — Helps you spot where AI fits in your own work, available in all six interface languages.
• Smoother course flow — Finishing a course now leads straight to the next one in your path, and end-of-course feedback is optional.

Fixes

Document and data source processing

• OneDrive sync no longer reprocesses files it already handled after an interrupted run, so syncs finish faster and stop re-indexing unchanged documents.
• After a failed database upgrade caused by a temporary lock, the vault now recovers on its own instead of staying unavailable until an administrator intervenes.
• Files stored in OneDrive whose names contain characters such as % or # can be re-read reliably; previously such a file could fail to load or be reported missing when the system needed to fetch it again.
• Data pools whose connection needs re-authorization are now retried on a slower schedule instead of every few minutes, reducing load and log noise; a manual sync is always available right away.
• Website crawls recover on their own after a database hiccup, instead of leaving that worker unable to crawl until restarted.
• Very large websites are now crawled to completion; the crawler continues where it left off instead of restarting from the beginning each time.
• If the folder a data pool is connected to gets deleted or renamed in OneDrive or SharePoint, the pool now clearly reports a configuration problem instead of showing successful syncs with no documents.
• If the folder a Google Drive data pool is connected to gets deleted, trashed, or unshared, the pool now clearly reports a configuration problem instead of showing successful syncs with no documents.
• Error details for documents that failed to process now name the affected file itself instead of an internal temporary filename.
• Documents that fail to process now always show the reason next to their status, and documents that recover no longer keep an outdated error message.
• Fixed OCR page processing in project tables adding credits to your balance instead of deducting them.

Other fixes

• Chats with the newer GPT-5 models no longer fail with an unsupported-effort error; the effort levels offered per model now match what the model actually supports.
• Editing a skill file and closing the editor now asks before discarding pending edits.
• AI Campus exercises now answer again for workspaces whose default model is set to automatic selection; failures now surface a message instead of failing silently.
• Creating a workflow with AI now shows each step as it's written instead of a spinner, and can be cancelled while it runs; the first step appears within seconds.
• Clearer team role management — Team roles are now set from a labelled dropdown in the member list instead of an unlabelled icon button, so it's clear who is a team admin and how to change it.

• New guides for planning your AI rollout: Three new guides walk you through planning an AI rollout, checking a vendor against § 203 StGB, and comparing platforms by criteria instead of demos — plus two interactive checks and a personal roadmap with real dates.
• Lower prices for GPT-5.6 Terra and Luna: following OpenAI's July 2026 price cut, both models are noticeably cheaper — Luna in particular now sits in our lowest cost tier.

Fixed

• If sending a magic link fails unexpectedly, the sign-in page now shows a clear error message instead of technical system output.
• Documents added to a large project become searchable much sooner: the index no longer works in short bursts with long pauses, but keeps going until it has caught up.
• The welcome screen now shows a model that's actually available in your workspace, instead of switching only after your first message when your saved model had been disabled.

New AI models: DeepSeek V4 Flash and V4 Pro

DeepSeek V4 Flash and V4 Pro are now available, bringing a million-token context window along with stronger reasoning and coding performance. DeepSeek R1 is deprecated in favor of the new models.

More improvements

• AI-generated content is now marked — As required under the EU AI Act, a notice now appears wherever AI-generated content is involved: throughout the web app, in the Outlook add-in, and spoken out loud at the start of every voice assistant call.
• Image-heavy chats stay fast — Chats with many generated or attached images stay responsive, and images are easier to review directly in the conversation.
• Documents can show their figures — The assistant can now show a drawing or photo found inside a document as part of its answer, instead of only describing the text around it.

Fixes

• Website sources set to keep only the main content now store just the article text of each page, not menus and footers. Search results from these sources are cleaner, and pages stop re-processing daily when only navigation or footer content changes.
• Search finds more of what it should: documents with an umlaut in the file name, a section that had failed to be prepared for search once, and content near the end of very long documents.