Version 2.28.0

Four new models

Four new models are available — administrators can enable them in the organization settings:

• Claude Opus 4.8 — Anthropic's latest Opus with a 1M token context, sharper agentic coding, and stronger reasoning. EU-hosted via Vertex AI.
• Gemini 3.5 Flash — combines pro-level reasoning with high speed and understands text, image, audio, and video input. Enabled automatically for organizations already running Gemini-family models.
• Mistral Medium 3.5 — latest generation of the EU-hosted Mistral Medium model.
• Kimi K2.6 — successor to Kimi K2.5, hosted on Azure AI Foundry (Sweden).

Pick connectors directly in chat

Connectors like Google Calendar, Microsoft Teams, or custom MCP servers can now be selected directly from the @ menu in the chat input. Each pick activates the connector and inserts an @-mention in the message — same pattern as assistants and data pools. Active connectors appear as chips below the input and can be removed there with one click.

Import memories from other AIs

Settings → Memory has a new import wizard. Paste a prompt into ChatGPT, Gemini, or Claude, paste the answer back into meinGPT, review the extracted memories, and apply with one click. Carrying work context over from another AI now takes a few seconds.

Fixes & Improvements

• Per-function permissions for connectors like Outlook or Teams are now reliably honored when the connector is activated directly from a chat.
• Steadier chat stream on brief disconnects — When the tab switches or the network blips, the response now picks up live immediately on reconnect instead of staying silent until the next larger step.
• No more duplicate messages when the document or canvas view opens during an in-flight response.
• Excel Agent entry in the sidebar — Icon is properly centered in the collapsed sidebar; the hint text now shows up directly below the row instead of far to the right.
• Very large documents stay searchable — Documents that exceed the chunk cap are now automatically indexed via keyword search (BM25) instead of failing. They remain searchable, just without semantic similarity matching.
• Documents with tokenizer control strings index reliably — Documents that contain strings like <|endoftext|> as normal text (e.g. API docs, AI transcripts) no longer fail during token counting.