cat changelog.md

What's new

Changes to Oshu, newest first: new features, prices and privacy.

  1. BYOK now costs 39 € a month

    The BYOK plan, where you bring your own Mistral key, now costs 39 € a month (net) instead of 79 €. It still includes 50 agents, unlimited messages and an unlimited library. Mistral bills you directly for chat replies and document reading (OCR).

    Existing BYOK subscriptions renew at the new price.

  2. Chats are deleted at Mistral after every reply

    Mistral keeps conversations made through its Agents API until the account is closed. Oshu now deletes each conversation at Mistral right after the reply, for every chat. Before, this happened only for private chats. An hourly cleanup catches deletes that failed and removes earlier conversations of agents that run on our Mistral account.

    Your conversation history in Oshu is not affected and still follows your retention setting. With your own Mistral key (BYOK), each new conversation is deleted from your Mistral account after the reply as well; conversations from before this change stay in your Mistral account until you delete them there.

    Web search, which is off unless you switch it on for an agent, runs at Mistral on Brave Search in the USA. The agent form, the privacy policy and the sub-processor list now say so. If you use web search, mention it in your own privacy notice: the privacy snippet in the agent's Embed tab has an optional sentence for it.

  3. Mistral Large 4 and GLM 5.3, model names in Quality

    On BYOK and Custom you can now pick Mistral Large 4 (public preview) and GLM 5.3 for an agent. Large 4 thinks when you switch reasoning on; GLM 5.3 always thinks.

    On the other plans, the Quality options now name the model behind each tier, for example "Premium · Mistral Medium 3.5".

  4. Show the assistant's work

    Replies that search your libraries can take a while. A new setting on the agent's Configuration tab, Show the assistant's work, decides what visitors see in the meantime:

    • Subtle (the default for new agents): one quiet line with the current step.
    • Tool steps: every step with its duration and search terms, folded into one line once the answer is there.
    • Thinking: also the model's reasoning and any code it ran. It needs a model that thinks.

    If nothing new arrives for 3 seconds, the widget says it is still working. The setting replaces the "Show reasoning to visitors" switch: agents that had it on keep showing their reasoning, and show code as well once you pick Thinking in the new setting. Sources are now listed under the answer, one per line, and follow the Show citations setting.

    In the conversation view you can open the steps of each reply, including the address and title of up to ten results per web search. Visitors never see those results.

  5. The assistant can use the page the visitor is on

    A new switch on the agent's Configuration tab, Use the current page, sends the page's address, title, first heading and breadcrumbs with each message, so the assistant can prefer material about that page. Query strings and fragments are removed from the address. If the address exactly matches a page in one of the agent's libraries, a short excerpt of that page is added.

    It is a hint, not a filter: the agent still answers from all its sources. The switch is off by default, and in private mode no page details are stored.

  6. Private mode: let your visitors choose

    Private mode now has four settings on the agent's Configuration tab: Off, Visitor's choice, starts off, Visitor's choice, starts on and Always on. With a visitor's choice, the widget shows a private-mode toggle and remembers the choice for the browser tab.

    In private mode the visitor's messages are not stored; the assistant's replies are. The assistant still keeps the context of the chat: the widget sends the earlier private messages along with each new one (in very long chats the oldest are left out). On the workspace page, private chats stay in the list of earlier chats and reopen from the visitor's device.

  7. Widget text stays readable on any colour

    For themes made from the four base colours in the theme editor, the widget now checks every text and icon colour against the background it is painted on and adjusts it where needed: 4.5:1 contrast for text and 3:1 for icons (WCAG AA). The colours you pick still fill the bubbles and the launcher exactly. Themes saved before the base colours existed look as they did.

    Timestamps, subtitles and labels are darker than before, citation chips and code blocks take a shade of the assistant bubble, and the theme editor warns when a colour you override falls below the limit.