What we shipped recently

New models

meinGPT Web for web search

Web search in chat now runs on meinGPT Web, our own web-search model, for fast, current, source-backed answers — hosted in the EU. Existing API integrations that use the Perplexity Sonar models keep working unchanged.

Two new Google models

Gemini 3.6 Flash is now available for fast everyday reasoning, alongside Gemini 3.5 Flash-Lite as a low-cost, high-speed, EU-hosted option. Workspace admins can switch both on in the model settings.

Improved Meetings

Chat with your meetings

Meetings are now a knowledge source in chat. Turn Meetings on for an assistant, or add it to a single conversation from the + or @ menu, and the assistant searches across your meeting transcripts and reads them when they're relevant to the question.

A clearer meetings list:

The meeting list is now search-first, with a single Filter popover for participant, date range, source, and status — and removable chips for the filters you've applied — instead of five always-visible dropdowns.

Fixed Bugs

• Scheduled meeting notetakers are now private to the colleague who scheduled them — other members can no longer view or cancel them.
• Image generation with Gemini 3 Pro Image now uses Google's final release of the model instead of the discontinued preview version. Workspaces with the model enabled were switched over automatically.
• Fixed the credit top-up history table showing an incorrect amount in the Credits column.
• Occasionally several meinGPT notetaker bots joined the same meeting — now only one joins.
• Meeting durations now display consistently (a 44-second clip shows as 0:44, not 44), alongside smaller polish in the meetings view.

For administrators

• Website crawler page limit is now a total, not a per-run quota: The configured page limit now caps the whole source. Scheduled re-crawls no longer keep adding pages beyond it; already-indexed pages still update, and removed pages are still cleaned up.
• Smarter change detection for crawled sites: Website crawler sources now detect real content changes instead of cosmetic page differences, so unchanged pages are no longer re-processed every day — keeping crawled knowledge fresh with far less background load and more stable search on large sources.
• Crawler sign-in works with more login systems: Website crawler sources that sign in now handle more login setups, including sites that redirect to a separate single sign-on page. Previously some login-protected sites could never be indexed.

Overhauled meeting view

The meeting list stays fast even with many recordings: meetings are grouped by time, show participants and a preview image at a glance, and can be filtered by source (Bot or Live), participant, or company. In the meeting view the recording plays at the top (video or audio), with summary and transcript side by side instead of behind tabs.

Recordings made by the meinGPT notetaker bot now name their speakers automatically instead of "Speaker 1 / Speaker 2". Names are only assigned when the match is confident; otherwise they stay editable as before. Calendar-scheduled meetings also record who was invited, so new meetings can be filtered by company and are linked to invited meinGPT users.

Meeting search in chat

The assistant can now search past meeting transcripts directly in chat: ask what was discussed, decided, or agreed — narrowed by participant, company, or date — and it finds the relevant meetings and answers from their transcripts. It only sees meetings the person asking owns or that were shared with them.

For administrators

• Invite links now work even when the organization has reached its seat limit: new people join as viewers instead of being turned away, matching how email invitations already behave.
• Google Drive data pools connected to a folder inside a Shared Drive now sync correctly. Previously such pools reported successful syncs but found no files; they pick up the folder's content automatically after the update, without reconnecting.
• Automatic credit top-up is paused for now while billing is being improved — it could not always produce a proper invoice. Top up manually, and set a low-balance warning to get an email before credits run out.
• The low-balance warning can now be set directly on the credits page — enter a threshold and save to get an email before the balance runs out.

Fixed

• Clearer web search answers: Web Search now produces clearer, cited answers with meinGPT models while continuing to use live web sources.
• Switching the AI model in the middle of a chat after a tool step (for example after a generated script) no longer causes an error — the conversation continues normally.
• Excel files saved by certain third-party spreadsheet apps no longer fail to be indexed in the data vault — their contents are now searchable like any other spreadsheet.
• Corrupted legacy Excel (.xls) files are now cleanly marked as unsupported in the data vault instead of being retried on every sync without a visible error.
• Files with square brackets in their names are now found reliably by the data vault file search.
• Files uploaded to a data pool while a synchronization run is in progress are no longer falsely reported as failed — uploads land reliably even when both happen at the same time.

What's new

A rebuilt prompt library

Saved prompts used to live in a single flat list — fine for a handful, unwieldy once a team relied on them. The prompt library is now organized around folders: keep a prompt private, share it with your whole workspace, or scope it to specific teams. Personal and shared prompts each get their own tab, with search and sorting to find the right one quickly.

More improvements

• Files from API automations: Make automations can now receive downloadable files created by sandbox-enabled assistants, so generated reports and processed uploads are easier to hand to the next step.
• Cleaner slide templates: Presentation templates can now mark slides as optional or fixed — unused optional layouts are dropped from the generated deck, and protected closing or legal slides are kept unchanged.

Fixed Bugs

• Copying an AI answer now copies only the answer — the model's thinking steps and tool details are left out.
• Scrolling up in a chat no longer jumps back down to the latest message on its own.
• Files uploaded to assistants through the API (for example via Make or n8n) are read reliably again, including large text and JSON attachments that previously failed with a "file not found" message.
• On phones you can reopen the navigation sidebar after closing it — a menu button now sits in the top-left corner.
• Sponsored voucher access now activates reliably however you sign in (Google, Microsoft, or magic link), even if you already started a trial — no more accidental trial in place of your sponsored access.
• Deleted teams no longer linger in the usage statistics, and a team's usage is no longer attributed to a member after they move to a different team.

For administrators

Control membership per verified domain

Until now, whether people on a verified email domain could be invited into other workspaces was effectively an all-or-nothing rule. Each verified domain now carries its own toggle under Access & Security. Newly verified domains are protected by default — people on them, and members of the owning workspace, can't be invited or added elsewhere until an administrator allows it. Domains that were already verified keep their current access, so nothing changes for existing setups until the toggle is switched.

Fixed

• File-heavy chat messages start more smoothly: attachments that need the sandbox are prepared automatically, reducing waits caused by manual upload steps.
• Workflow variables written with spaces inside the braces (like {{ name }}) are filled in correctly again, so the assistant no longer asks again for information that was already entered.
• Web research is more reliable: temporary search-provider errors and timeouts are retried automatically instead of failing the request, so research no longer stops early with an incomplete answer.
• Downloading or editing an AI-generated file (such as an Excel export) from older chat history works again — the download link no longer expires after a while.

GPT 5.6

Three new OpenAI models join meinGPT, all hosted in the EU, each a clear capability tier:

• Sol — the flagship, for the most demanding professional work: complex reasoning, coding, and agentic tasks where quality matters most.
• Terra — the balanced everyday model for general knowledge work, at a lower cost than the flagship.
• Luna — the fastest and most affordable of the three, built for high-volume work while keeping strong reasoning.

The GPT-5.6 family is built around results per euro rather than a lower sticker price: the models reach comparable or better answers using far fewer output tokens, so the cost of a completed task tends to drop versus the previous generation — most sharply on Terra and Luna. The models are available now; administrators manage which models a workspace can use in the settings.

What's new

Web search and Deep Research, rebuilt

Web search and Deep Research now run on a new search service behind the scenes and come with a refreshed interface. The biggest change is for privacy: searches now handle personal and sensitive data more safely, so the old "no sensitive data" restriction is gone. Administrators can turn the web search app on or off for their workspace.

For administrators

Switch users between Member and Observer via the API

The User Management API can now change a user's role between Member (uses a seat) and Observer (no seat) programmatically — so you can automatically free up a seat when someone finishes and reassign it when they return. Seat limits and billing stay in sync, exactly as they do in the interface.

Fixed

• More reliable syncs from OneDrive, SharePoint, and Google Drive: when a provider asks to slow down, all transfers now pause together instead of failing, so large syncs finish without dropped files.
• Deep research in the chat now stays within the approved sources defined in an assistant's instructions, even for broader, multi-step research questions.
• Workshop invitations and notifications now use a consistent, personal tone. The cancel and update dialogs clearly state what changed, the preview reflects the entered reason before you send, and trainers are notified by email when their workshop role changes.
• Deleting a workflow step no longer leaves behind a leftover variable that couldn't be removed.

New features

Smarter document attachments

Attached documents are handled more efficiently. Text from PDFs, Word, and PowerPoint files is read directly, and many or very large files no longer overload the context. The result: answers that draw on several attachments at once become noticeably more reliable.

Connectors

This release is a big step for connectors — broader reach, tighter permissions, and finer control over what each assistant can touch.

• Official Asana & HubSpot integrations Asana and HubSpot now run on their vendors' official servers. Asana gains project creation, status updates, portfolios, teams/users, and workspace-wide search; HubSpot moves to secure OAuth instead of a manually created private-app token. Existing assistants and chats are migrated automatically — you only need to reconnect once.
• Outlook shared mailboxes & advanced actions Outlook assistants can now work with shared and delegated mailboxes, not just your own inbox — so a team assistant can read and reply from a shared address. A set of advanced email actions comes along with it. The extra Microsoft permissions these need are requested only when you turn the actions on, never at initial connect, and deleting emails is now opt-in per assistant rather than granted by default.
• Least-privilege permissions Connectors now request only the permissions the enabled functions actually need — Gmail read-only no longer grants write access, for example. Turning on more functions later prompts for the missing permissions before saving, so a connection never carries more access than it uses.
• Folder-scoped assistants Assistants can be restricted to specific folders (Google Drive, OneDrive) or sites (SharePoint) per assistant — so a single assistant reaches exactly the material it should. The personal restriction on the connection always remains the upper bound.

For administrators

A full settings audit log

Every security- and privacy-relevant settings change is now recorded: who changed what, when, and why — including changes made by meinGPT Support. Filter by category, impact, person, and date, and open any entry for the exact before-to-after detail. It gives admins a clear, reviewable trail for compliance and internal accountability.

Share custom AI apps precisely

Custom AI apps were previously all-or-nothing across the workspace. They can now be shared exactly where they belong: with everyone, with admins only, or with specific teams and people. Members see exactly the apps meant for them — nothing more.

Data Vault health alerts

meinGPT now emails workspace admins when a data vault has trouble — a connected source losing its connection, the vault being unreachable, or a sync repeatedly failing. Choose which alerts you receive under Settings → Notifications.

More improvements

• Preview files directly in chat — Sandbox files such as code, Markdown, PDFs, Word documents, CSV files, and images can now be checked before download.
• More flexible Code Sandbox — The sandbox now accepts more input variations: runtime names like "typescript"/"py", an omitted language (defaults to Python), absolute paths, and single documents. Permanent errors like "file not found" are no longer retried needlessly, so runs abort less often.
• Updated image model lineup — The Recraft V3 and Recraft V3 SVG models — and with them vector (SVG) generation — have been retired. Workspaces that relied only on Recraft now default to FLUX.2 Pro.

Fixed

• Data pool search now recovers on its own after brief server hiccups instead of staying unavailable until a manual restart.
• Website data pools sync more reliably: successfully completed website crawls are no longer incorrectly reported as failed.
• Data Vault searches no longer fail when a query contains invisible control characters — for example in text pasted from PDFs. Such characters are now removed automatically.
• Markdown files with deeply nested lists are now indexed reliably during Data Vault import instead of being dropped from search.
• Calendar events and emails from connected Google and Microsoft accounts now show dates and times in the local timezone instead of UTC.
• Creating and editing Outlook email drafts via the AI tools works again — previously these actions failed as soon as a subject or body was set.
• Long prompts from the prompt library are fully usable again: the prompt text scrolls inside the dialog and the "Use prompt" button always stays visible.
• Workflow steps marked as skippable are again correctly skipped at run time as soon as their condition is met.

Claude Sonnet 5

Claude Sonnet 5, a new Anthropic model, is now available – hosted in the EU via Google Vertex, with a 1M token context window and a balance of speed and quality. It is well suited for coding, analysis and agentic workflows. Admins can enable the model in the organization settings; it is switched on automatically for organizations that already have a Sonnet model active.

Further improvements

Internal maintenance and stability improvements with no visible changes.

Internal migration

Internal maintenance and migration work with no visible changes for you.

A redesigned entry experience

The sign-in and sign-up pages and the workspace selection have been redesigned: a cleaner, more modern look for a clearer start. You can now filter the workspace selection, and longer lists load smoothly as you scroll.

Choose the thinking effort per model

You can now set how hard a model thinks — from Minimal to High — right in the chat, without switching models. More effort gives more thorough answers but takes longer and uses more credits. Your choice is remembered per model.

Redesigned assistants overview

The assistants overview has been redesigned: clearer cards, filtering by department, an examples gallery, and a tidier action menu per assistant. You'll find the right assistant faster.

Jira and Confluence in the new Atlassian connector

Jira and Confluence are now combined into a single "Atlassian" connector. Existing assistants are migrated automatically — you only need to reconnect once. Note: Atlassian Cloud is required (no Data Center/Server).

Create meetings faster

Creating a meeting is now faster and clearer: one click on "New recording" shows every option right away — audio/video, transcript, and live recording. You get straight to the one you need.

Improved PowerPoint import and slide editor

The PowerPoint import and slide editor have improved: correct font size and weight (including Light), an alignment control, live formatting changes, tables that fit with their borders, and a slide's purpose (e.g. a chart) recognised when filling the template.

For administrators

• Tool Management is now enforced — tools you can switch off (image generator, documents, code sandbox, and slides) now actually stay off. Previously they were still used despite being disabled.
• Higher per-IP API rate limit — the per-IP API rate limit has been raised substantially. Organizations that reach the API through a shared (NAT) IP address are no longer throttled collectively by mistake.

Fixes & Improvements

• Claude Opus 4.7 and 4.8 available again — requests to these models via the API and in chat work again; previously they failed with an error.
• Voice input reliable again — cancelled or never-started microphone recordings release their slot immediately, so the "could not start voice session" error appears far less often.
• Tools reliable again in chats with history — on Claude and Gemini models, tool calls in chats with prior tool history (e.g. the Code Sandbox) now receive their arguments correctly again.
• Help for chat errors — every chat error now has a "Why am I seeing this error?" link that opens a help page with the background and fix.
• Default-model selector loads reliably — the default-model selector in settings no longer briefly appears empty.

Multi-file projects in the Code Sandbox

The Code Sandbox now supports projects with subdirectories (e.g. src/app.js) and dotfiles like .env. At the same time, failed or aborted code runs happen far less often — building multi-file projects is faster and more reliable.

Specific messages for chat errors

For common errors, chat now shows specific, actionable messages — out of credits, a premium model needing credits, too many or oversized images, or an unsupported file type — instead of a generic notice. Credit errors link admins straight to billing settings, while members are pointed to a workspace admin.

For administrators

In Mandatory Trainings settings, the “skip lesson” and “show hint” buttons (including the sample solution) can now be turned off for courses, so learners complete the exercises themselves.

Fixes & Improvements

• No more internal author notes in course text — internal notes from course authors are no longer accidentally shown in the course text.
• Stable file search with date filters — file searches in data pools with created or modified date filters are stable again. The filters apply only to file metadata, not to years in the search term.
• Faster browsing of large knowledge libraries — browsing large knowledge libraries is faster, and folders deep inside very large libraries now appear reliably when you navigate into them.