This is a preview page for the upcoming 2025.52 release to announce the necessary infrastructure changes early before the release as can be seen in production.

Release Overview

• 2025.52

• Unique hosted: January 7, 2026

• Self hosted: January 9, 2026

Chat Platform

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

User Model Usage Analytics: Model usage logs can now be requested and downloaded from the analytics page

Admin Code Execution In-Chat Upload: Code Execution will now by default upload files generated by the code execution agent to the chat instead of loading files to the knowledge base. Files will still be scanned for malware; however, this provides improved data privacy limiting access to the individual who created the chat. Reverse compatibility exists allowing admins to still opt for uploading code execution artifacts to the Knowledge Base.

🔧 Improvements

User PDF Viewer: Improved PDF viewer scaling and display for landscape-oriented documents like PowerPoint presentations.

🐞 Bug Fixes

User Performance issues: Reduced excessive CPU consumption during scrolling in the Chat Platform, improving overall responsiveness.

Agentic Table

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

User Agentic Table Metadata EXPERIMENTAL: Metadata and table instructions can now be defined at the sheet level, allowing answers and knowledge to be tagged with contextual information such as clients, funds, or time periods. This improves retrieval accuracy by enabling metadata-aware filtering in search and the answer library. Existing sheets without metadata continue to work unchanged.

Rows ingested into an Answer Library table inherit table metadata. The resulting knowledge base file will be tagged with the table metadata and a json representation of the metadata will be included in the embedded text.

🔧 Improvements

User AG Grid Library Upgrade: The Agentic Table now uses AG Grid version 35, bringing updated functionality, performance improvements, and fixes while ensuring all existing features continue to work as expected.

User Streamlined Button Styling and Wording: Buttons in the Agentic Table onboarding and the Share action now use a more subtle variant with improved wording, creating a more balanced and consistent experience.

🐞 Bug Fixes

User Updated Error Message Wording Error messages have been reviewed and updated to consistently reference the Agentic Table, removing outdated Magic Table terminology.

Unique AI

🔧 Improvements

Admin Custom Language Model Metadata via Environment Variables - Assistant Core now supports loading language model metadata from an environment variable, allowing clients to define and expose custom language model info in a convenient and standardised way. Model definitions provided via LANGUAGE_MODEL_INFOS are consumed exactly like the predefined models bundled with the unique-toolkit, including capabilities and token limits. This enables seamless integration of custom and provider-specific models. See here for more details.

Admin New approved models in dropdown for UniqueAI

• GPT 5.2 has been added following a comprehensive quality evaluation, showing a strong balance between performance, speed, and response clarity. Compared to GPT 5, it delivers a slightly faster time to streaming, improved output formatting, and highly consistent behaviour.

• Gemini 3 Flash Preview was added after achieving the highest overall evaluation scores, with best-in-class performance across both response quality and time-to-stream speed. The model produces the most refined and structured outputs and provides richer contextual explanations, though key information may occasionally appear deeper in the response rather than immediately upfront.

Note: Both models (GPT-5.2 and Gemini-3-Flash) are only available with a global deployment model. A restriction to Europe/Switzerland is not yet available.

🐞 Bug Fixes

Admin Improved Qwen Model Compatibility for Forced Tool Calls - A bug has been resolved where forced tool calls were not working correctly in UniqueAI when using Qwen 3 models. Due to a different convention in how Qwen models handle forced tool calls, the expected behavior was not triggered reliably.

Language Model

New Model Deployments:

Anthropic Claude Opus 4.5

Admin Claude Opus 4.5 is available through LiteLLM using the following enum:

• litellm:anthropic-claude-opus-4-5