Release Notes

Skills only get the connected servers they name. A skill no longer receives every connected server in its space automatically. Platform tools and knowledge remain available by default, but a connected server is offered to a skill only when the skill lists it among its allowed tools; a wildcard such as mcp__crm__* takes every tool of a server named crm. If one of your skills uses a connected server without naming it, add that server to its allowed tools. This narrows the default described in the 21 August release.

What's new, inside Aimable. A card at the bottom of the sidebar, in both the Workbench and the Console, shows the headline of the latest release. Open it for the full release history with filters, mark the changes that help you, and send us feedback with attachments. You can always get back to it from the user menu under What's new & feedback.

Updates no longer interrupt your work. We now start each new version alongside the running one and switch over once it is healthy, while the old version finishes what it was doing before it stops. An answer that is being written or a skill run in progress carries on through an update, and the brief outage that used to come with every update is gone.

Side chats. Ask something on the side without cluttering the conversation. A side chat opens in a panel next to the thread, from the button in the chat header or by selecting text in an answer. It reads the conversation it sits beside, but nothing you ask there reaches the main thread unless you take an answer over. Side chats are saved, private to you, and left out when you share the conversation.

Queue your next question. Type a follow-up while an answer is still being written and it waits above the composer instead of going nowhere. Queued messages go out together when the answer finishes. Edit one before it is sent, or push it through to stop the current answer and send it right away; the partial answer stays in the conversation, marked as interrupted.

Open a source next to the conversation. Clicking a citation now opens the document in a resizable pane beside the chat, scrolled to the cited passage and highlighted, instead of downloading the file. This works for PowerPoint files too. Web sources still open in a new tab, and a file type that cannot be previewed says so instead of failing.

Collections in the Workbench. Each space now has a Collections page showing the knowledge it draws on, both its own collections and those shared with other spaces, with file counts, processing status and sharing at a glance. Space curators can create collections, add existing ones, rename or delete them, and upload by dragging files in, with progress per file and a plain reason when a file fails. Everyone else sees the page read-only.

Uploads that don't hold you up. Attaching a file is instant: processing continues in the background, the file shows real progress (page 12 of 48, for instance), and you can keep typing. Send while files are still being processed and the message goes out as soon as they are ready. Several files are processed at once, and a file that fails no longer sinks the rest; it drops out with the reason and the message goes ahead. With a model that can read images, scanned PDFs don't make you wait at all.

Attach emails. Outlook .msg files and .eml files can be added to a message and are read like any other document. A file that cannot be attached now tells you why, and that it won't be sent, instead of looking attached when it isn't.

Spreadsheets are analysed, not skimmed. Ask about an Excel or CSV file and Aimable writes and runs code against the actual workbook, rather than answering from fragments of it. In collections, each sheet is indexed as one summary of its headers, sample rows and dimensions, so spreadsheets no longer crowd your documents out of search results. Files produced along the way are listed under the answer, ready to download.

Edit images in the conversation. In spaces with an image model, ask Aimable to change an image it generated, or upload a picture of your own as the starting point. When the image provider turns a request down because the account is out of credit, you are told so instead of getting a vague error.

Images from connected servers appear in the chat. When a tool on a connected server returns an image, such as a chart, a screenshot or a diagram, it now shows inline in the answer. Click it to open the original.

Download a conversation. Save a whole conversation as Markdown or Word from the conversation menu or the chat list. The menu of an open conversation can now also move it to a folder.

Skill runs you can build on. Ask a skill to adjust what it just produced and the new run carries on from the previous run's work in the same conversation, so a small change is a quick revision instead of a full rebuild. The run card has a labelled Results block, downloads carry the time of the run in their name, and Aimable knows where a run's output is when you ask for it. A run that fails says why, and a skill started from a conversation only sees the branch you are on.

Know when a model is down. The model picker marks a model that is currently failing, for example because its key was rejected or its credit ran out, so you can choose another before you start instead of finding out halfway through.

Privacy Check is back in the composer. The button that reviews personal data before a message is sent sits on the composer toolbar again rather than in the settings menu, under one name throughout the Workbench: Privacy Check.

Run your organization's skills on the open workbook. In the Excel add-in, a skill now works on the workbook you have open. The add-in hands it the file, lets you answer the skill's questions before and during the run, shows the steps as they happen and lets you cancel. New sheets the skill produces are added to your workbook, and a result too large to insert smoothly is offered as a download instead, so Excel stays responsive. See Using the add-in in Excel.

The add-in can see your sheet on any model. When the model you work with cannot read images, the add-in has a vision model read screenshots of your sheet for it, captured sharp enough to tell the cells apart, so questions about how a sheet looks work whichever model is selected.

A tidier add-in. Rename a chat from the chat list, and new chats name themselves after the first exchange. Edit a prompt you already sent and run it again. The chat list shows when each chat was last active, the model's thinking folds into a single row, and a running tool shows how long it has been going.

Attachments and knowledge work together. Attaching a file no longer stops Aimable from searching your collections for the rest of the conversation, and deselecting a collection takes effect on the message you send. When Aimable reads a whole file instead of searching it, the answer still cites it. An image in your message no longer makes Aimable lose track of your other attachments, editing a message keeps the files you uploaded with it, scanned pages inside an otherwise digital PDF are now read, and macro-enabled workbooks (.xlsm) work with the spreadsheet tools.

The Office add-in stays on track. A slow start to a long answer is no longer mistaken for a dropped connection, retries are visible, and when the model is still busy after writing text you see "Still working" with a timer. PDF tools work reliably, a crash when the model reused a tool-call identifier is gone, a write to a range of the wrong shape is refused with a clear message, a failed skill run shows the reason, and a macro-enabled result that cannot be inserted comes with an explanation.

Smaller annoyances. Switching space takes one click again, typing a name in the share dialog no longer blanks the page, and a refused action in a space tells you which permission is missing. Conversation titles are no longer lost when the title model is busy. Servers that connect on behalf of the whole organization no longer show up in your personal Connections as a button that always fails, and finished files no longer count as processing in a collection. Privacy Check no longer marks bare small numbers as personal data, and a text selection survives while an answer is still streaming in.

Bring your own model keys, end to end. Adding a provider key no longer dead-ends. Every key picker ends in an option to add a key, and creating one switches the model on straight away. When you add models to a space, the catalog shows the models that have no key yet and activates one for the organization and the space in a single step. A new key is checked with the provider before it is saved, with the provider's own reason if it is rejected, and only model keys are offered where a model key is expected. See LLM providers.

See which models and keys actually work. The Model Catalog gains a status column showing, for every activated model, whether it works or why it doesn't: key rejected, model gone, out of credit or call failed, with the provider's own message and the time of the last check. Check now runs the same checks on demand, the admin dashboard flags anything broken at the top, and tenant administrators are notified when something stops working.

Costs in euros, at the rate of the day. Every cost in the Console is now shown in euros, converted at the European Central Bank rate for the day the cost was incurred, history included. Dashboards and traces now show the same amount for the same spend, and the cost cards state which rate they use. See Observability.

Token usage next to cost. The tenant dashboard shows total tokens alongside interactions and cost, with a breakdown by space and a tokens column in the per-space table. Models no longer go missing from the token usage by model chart.

Long conversations cost less. Prompt caching is now on by default in conversations, so on models that support it a long conversation no longer pays full price for the same earlier context on every turn.

New models in the catalog. Added this cycle, through Tensorix: GLM 5.3, GLM 5.3 Flash and Qwen3.8 Flash Next, with a choice of reasoning effort. The models that power background tools such as image generation now sit in a clearly named Tool & background models section of the model policy, instead of behind an unlabelled gear icon.

Decide what a space shows. A new Visibility section in a space's governance settings lets you hide the model picker when the space offers a single model, hide the navigation menu so the space shows only chat, and turn off sharing conversations from the space. Source citations can be hidden per space, and the Workbench Collections page can be switched off for a space; it is on by default. See Governance.

Programmatic access in one flow. Giving an API key access to a space used to take five steps across three screens. A Programmatic access wizard on the Principals page now creates the principal, its key and its role in the space in one go, with presets that make least privilege the easy choice: space automation, full space admin or read-only. The API keys page lists every key in the organization with its owner and status, expired keys included, and says why when it cannot load. API principals can be renamed, and the Principals list can be filtered by status.

An icon for your organization. Give your organization its own icon in Global Settings, chosen from the same icons and emoji as spaces. It appears in the organization switcher, so people who belong to several organizations can tell them apart at a glance.

Sign in with Salesforce. Salesforce joins the supported identity providers. Enter your Salesforce My Domain URL in the identity provider settings and Aimable derives the rest; sandbox domains work too. See Identity providers.

Skills written for Claude run as written. Skills authored with Claude expect the built-in Word, Excel, PowerPoint and PDF skills and a working LibreOffice. Aimable's sandbox now provides the same environment, so a document-building skill written that way works when you upload it, without rewriting it and without the skill spending its time working around missing tools. See Building skills.

Connect servers without registering an OAuth app first. For connected servers that support dynamic client registration, GitHub among them, Aimable now registers itself as the OAuth client and reuses that registration, so nobody has to create an app at the provider before the server can be connected.

Connectors and connected servers. Editing a connected server keeps its saved OAuth settings and secrets, and a stale API key can no longer override its OAuth client secret. Google Drive now lists the shared drives you are a member of, which used to need Google Workspace administrator rights, and folders with more than a hundred files sync completely. A connector whose sync died is no longer stuck for good, and long syncs no longer fail partway through.

Skill imports. A skill bundle is no longer rejected over a stray file next to the skill folder, such as a macOS .DS_Store, or over the capitalization of its SKILL.md, and when an import is rejected the Console tells you why.

More complete cost figures. Some background model calls, including the ones behind Privacy Check, are now recorded and priced; usage logged under a bare model name is attributed to the right API key; and Cost by Model no longer splits one model over two rows.

Start a skill run with files attached. POST /v1/spaces/{space}/skills/{slug}/executions/upload accepts the same inputs as the JSON route, plus files[], as multipart. The files are placed in the run's workspace under inputs/materials/ for that run only; nothing is indexed into a collection.

Admin API additions. GET /v1/admin/api-keys lists every key in the organization. PATCH /v1/admin/principals/{id} accepts a permissions list, so an API principal can be granted individual permissions on its tenant membership without a whole role, and API principals can be renamed. The principals listing takes a status filter, and POST /v1/admin/secrets/validate checks a provider key with the provider before you store it, without storing or logging the key.

Skill runtime and manifest. A skill's model_hints can set its own reasoning_effort, and min_context is honoured when a model is chosen for a run. web.fetch supports POST for public search APIs. Tool results are capped per call with a footer stating the full size, and workspace.fs.read pages through large files with offset and limit.

Completions built for long sessions. /v1/chat/completions sends keepalive comments while the model is still working and applies first-token and idle timeouts, so a slow start is not cut off along the way and a dead upstream ends in an error instead of a hang. Prompt caching is on by default (send X-Prompt-Cache: false to opt out), the system prompt carries the date but not the time, reasoning from earlier turns is no longer sent again, and history compaction falls back to the latest turn boundary when there is no cleaner place to cut.

Tighter security. Requests that connectors and web tools make on your behalf can no longer reach private or internal network addresses. Aimable only answers to the addresses a deployment actually serves, organizations are kept more strictly apart, and deleting a file removes exactly that file's content from search: all of it, and nothing else.

Reliability. A skill that is waiting for your answer no longer fails after two minutes. A spreadsheet formatted across its whole grid can no longer exhaust the server's memory when it is previewed, a skill run that produces nothing now ends with an error instead of an empty result, bursts of image analysis back off instead of failing, and the skill version offered in a conversation is the version that runs.

About the beta features in this release. Items marked (beta) are built and running, but we roll them out per organization rather than switching them on everywhere at once. If something below sounds useful to your team, get in touch and we will enable it for you and take you through it.

Chat with an agent (beta). Pick an agent from the composer and the whole conversation runs as that agent, with its own instructions, its own model and its own set of tools. The header and the individual messages show which agent is speaking, so a conversation involving several agents stays easy to follow.

Agents that work together (beta). An agent can hand work to other agents in your space: consult one specialist and wait for the answer, ask several in parallel, or send work off in the background and have it report back into the conversation when it lands. Every hand-off appears as its own card that you can open to see the exact question, the answer and anything that went wrong. When a specialist needs a clarification, the agent coordinating the work answers it where it can, and only comes to you when it genuinely needs a human decision.

Agents that know their colleagues (beta). An agent now knows which other agents are available to you and points you to the right one instead of guessing outside its own expertise. Each agent has its own icon and appears by name on the space's Members page, and a one-to-one conversation starts on the model configured for that agent rather than the space default.

Chat folders and archive. Organize your conversations into personal folders, and archive the ones you are finished with.

Find and pin your conversations. Press Cmd+K to open a search palette and jump straight to any conversation, and pin the ones you keep coming back to so they stay at hand.

Reply to a specific passage. Select part of an answer and quote-reply to exactly that passage, instead of describing which bit you meant.

Editing a message creates a real branch. Editing any message — including the very first one — forks the conversation into a branch that survives a reload, and you come back to the branch you were reading. Explore an alternative without losing the original thread.

Leaving a conversation no longer costs you the answer. A response that is still being written keeps running when you navigate away and picks up where it was when you return. While a new conversation is still being named, the sidebar shows an animated indicator instead of a blank entry.

Conversation titles stay in the right script. An automatic title no longer drifts into an unrelated script — a Chinese title on a language-neutral first message, for instance — when there is little to go on.

Skills can write and run their own code (beta). A skill can now write code and run it in a sandboxed workspace — analysing a spreadsheet, reshaping data, producing a file — with a broad set of libraries available, a disk limit per workspace and a cap on how many runs happen at once. Sandboxed code never has network access: a skill that needs to fetch something uses the fetch tool, which writes the result into the workspace.

Skills start with everything the space offers. A skill can use the tools, connections and knowledge available in its space by default, with its allowed-tools list acting as a restriction rather than a list you have to build up from nothing. Files and results from the current conversation — and from earlier turns in the same thread — are staged into the skill's workspace, so it works from what you already shared instead of asking again.

A smoother skill runner. Drop files straight onto a skill's upload block, reach skills that present their own interface from a launcher in the sidebar, and download what a skill produced on a canvas.

Aimable can remember what matters (beta). Personal and shared memory keeps facts and context across conversations, so you stop re-explaining how your team works. Everything remembered is visible and editable, and a note appears under an answer when something new was learned.

Skills that run themselves (beta). A skill can run on a schedule, or when something happens in a space — a report every Monday morning, a summary whenever a document lands. An Automations overview shows what is scheduled, what ran and what it produced, with notifications when a run needs you.

A Console link in the user menu. Administrators reach the Console directly instead of retyping the address.

Clearer errors, in English. Tool, connector and authorization failures now surface in the conversation instead of the answer stalling silently, posting in a space you are not a member of explains what to do about it, and the remaining Dutch error messages have been translated.

Canvas polish. Copying from a canvas — including from the edit view — gives you real Markdown, and copying a code block no longer drags the fence markers along. You can scroll a canvas while it is being written, the formatting toolbar stays in view in long documents, task-list checkboxes render as checkboxes, and closing a canvas gives the conversation its full width back.

Working with images and personal data. Pasting a screenshot no longer fails on an unrecognized detector type, redacting several images in one message is queued instead of run all at once — which removes the crashes and failed uploads that came with it — and the entities dialog keeps its size while you add, remove and search entries.

Attachments stay where you put them. A file attached to one message is no longer silently re-attached to every message after it.

Smaller annoyances. Opening a conversation no longer bumps it to the top of the sidebar, the composer no longer covers the last lines of an answer, and opening a thread straight after switching space loads the right conversation.

See which parts of the product cost you money. The tenant and space dashboards gained a breakdown of spend by surface — chat, skills, agents, meetings — alongside the existing per-model and per-key views, so you can see where the bill comes from rather than only which model produced it. Latency and error-rate tiles have made way for it, and every chart now shares one palette, so a given space or model keeps the same colour wherever it appears.

Spend caps and budget alerts. Set a spend cap per space and for the organization as a whole, choose the period it resets on, and define the percentages you want to be warned at — for example at 50%, 80% and 100% of the cap. A budget card on both dashboards shows spend against the cap as it accrues, and alerts go out automatically when a threshold is passed, so a runaway workload surfaces while you can still act on it.

A much bigger choice of icons. The icon picker now searches the complete Lucide icon set — around 1,760 icons — plus emoji, in one Slack-style search box. Give every space and every agent an icon that actually reflects what it is for, so people recognize the right one at a glance instead of reading a list of similar names. Picked something you would rather undo? An icon can be cleared back to the default.

New models in the catalog. We keep the catalog current with what the labs actually release. Added this cycle: Claude Opus 5, GPT-5.5, GPT-5.6 (Sol), GPT-5.6 Luna, Gemini 3.5 Flash, Kimi K3, DeepSeek V4 Flash and the new GPT Image 2 and GPT Image 1 Mini image models. They join a catalog that already spans the Claude family (Opus, Sonnet, Haiku and Fable), the GPT-5 range, Gemini, GLM, Kimi, Qwen, Grok, Llama, Mistral and DeepSeek — across direct, Azure Foundry, Scaleway and Tensorix routes. Superseded versions were retired in the same pass so the picker stays honest.

Manage agents per space (beta). Agents have their own page under a space, with the agent list and the delegation governance settings side by side. Set an agent's instructions, choose its model, restrict it to specific collections, and decide which surfaces it appears on. The allowed-tools picker is fed by the live tool registry, so a space's own connected servers and skills are offered as tools straight away. Agent names are unique per space rather than per organization, so two spaces can each have their own Writer.

Assign permissions in bulk. The role editor takes several permissions at once instead of one at a time, and the full permission set is now offered.

Connector setup tells you the callback URL. The connector dialog shows the OAuth callback URL you need to register, with a link to the provider's registration page, so setting up an integration stops being guesswork.

Pin the Office add-in to a space. Choose centrally which space the Word and Excel add-in works in, instead of asking every user to select one.

Tune memory per space (beta). Memory is configured from the Console, for the organization as a whole and per space: switch the reflector on or off, decide whether semantic recall is used, set how many past episodes may reach a single turn, choose whether proposed memories apply automatically or wait for review, and point it at a lighter auxiliary model. An insights page shows what memory is actually doing — writes, additions, deletions, evictions over the last seven days, the median age of what is stored, and the proposals still open — broken down by scope and space. People can review and export their own memory from their settings.

Administer automations (beta). Switch automations on per space and cap how many a space may hold. Build a schedule with a cron builder or subscribe to an event, set the retry policy, and follow what happened on a detail page with full run history and per-run status. Viewing automations and managing them are separate permissions, so a team can watch what runs without being able to change it.

A clearer space settings area. Space settings moved to a sidebar submenu structure, with the occasional settings gathered under Advanced tools on the Governance page — so what you touch weekly stops competing with what you touch twice a year.

Roundtable configuration (beta). Both Roundtable modes are toggled from the tenant Features dialog. Templates take a manual sort order, so your own templates lead the picker instead of the built-in examples, and an agent's model is chosen from the models its space actually allows, with the session's writer marked as protected.

Model catalog hygiene. Models you deactivate in the catalog no longer appear in the space model pickers, the separate title-generation model choice has been removed in favour of a fixed lightweight model, and the meeting-summary override only appears when meeting capture is switched on.

Agent changes save reliably (beta). A deleted agent no longer keeps its name reserved, editing only an agent's instructions no longer rewrites its tool list, and saving an agent no longer fails validation against a backend that predates its newest settings.

Space-scoped grants work across the board. The agent endpoints, the execution routes including responses, and agent.invoke all accept a space-scoped grant instead of requiring the tenant-admin role, so an API key can be given exactly the space it needs and nothing more.

Agent API additions (beta). Agents carry available_in_chat and available_for_delegation to control where they appear, an allowed_collections scope, and an allowed_tools list that is now enforced server-side at the point tools are offered to the model — an agent with a restricted list still inherits the connected servers and skills of its space.

Delegation tools (beta). delegate_invoke_agent, delegate_fan_out, delegate_invoke_agent_async and delegate_answer_clarification, enabled per space. Fan-out requires a declared dependency analysis marking the tasks independent; a declared dependency is rejected with instructions to retry sequentially, so dependent work cannot be silently parallelized. A delegated run is an ordinary skill execution, so it inherits the existing audit trail, cost accounting and tracing.

Trigger API (beta). Cron triggers take a five-field expression and an IANA timezone with a 60-second minimum, and do not catch up after downtime — one fire, then the next future slot. Event triggers subscribe to a platform event type narrowed by equality filters. Retries and dead-lettering have separate budgets, and a trigger that keeps failing switches itself off. Per-space cap and feature flag, tenant-admin only.

Stepped history compaction. Long conversations compact from 200 to 100 messages, so extended add-in and chat sessions stop re-billing the full context on every turn.

Space-scoped screenshot-analysis endpoint. Backs image understanding for models without native vision support.

Per-organization model base URLs are honoured. Model invocation now reads the configured base URL for a tenant's model, which fixes authorization failures on Azure Foundry-backed models across the Responses, agent and meeting surfaces.

Errors reach the caller. Tool and connector errors — including connected-server and authorization failures — are emitted on the streaming error channel instead of the stream ending silently, validation errors carry the actual validation detail, and a turn that produces no output raises an explicit error rather than returning a blank answer.

Accurate cost attribution. The price catalog now handles both bare and vendor-prefixed model names, closing gaps where the same work was counted at zero or counted twice. Calls made on your behalf by agents and by memory are attributed and priced alongside the rest.

Connected servers and skills in Word and Excel. The Office add-in now reaches the same connected servers and skills as the Workbench, through the same tool endpoints. That means the add-in can pull from your systems and run your organization's skills without leaving the document you are working in — drafting from live data in Word, or having a skill do the analysis in Excel.

A more workable Office add-in. The prompt box can be dragged to the size you need, so writing a longer instruction no longer means typing into a letterbox, and the add-in follows the space your administrator pinned instead of asking each user to choose one.

Meeting notes that land in your knowledge (beta). Aimable Capture joins your meeting, transcribes it, and writes the notes into a collection — so the moment the meeting ends you can chat about what was said, ask follow-up questions, and put agents to work on the outcome. No more notes that live in someone's document and are never opened again.

Transcription that knows your room (beta). The recognizer is biased toward the names of the people actually in the meeting, your custom vocabulary is applied to the live transcript as well as the final one, and you can select a word in either and correct it on the spot.

Meeting capture controls (beta). Organizations get a sensible default limit on simultaneous recordings and a clear message when it is reached, and a tenant administrator can force-stop a recording that is stuck or abandoned so it stops consuming the quota. A meeting can also be created ad hoc for a single session.

Aimable Roundtable: a meeting that writes its own document (beta). Agents join your meeting, listen, and build the document with you on a shared canvas — an outline that fills in as you talk, research dossiers, and a live feed of what each agent is doing. Ask for research out loud and an agent picks it up, shows the request straight away, and comes back with a dossier, spreading the work across specialists where that helps.

A Roundtable you can direct (beta). The writer works from the outline, merges superseded action items instead of stacking duplicates, protects your manual edits from being overwritten, and follows the document's language rather than whatever was last spoken. Agents respond when genuinely addressed by name, respecting whether each one is listening or muted, and tolerating the mishearings that come with live speech. Sessions run from per-space templates with a designated moderator, and can be downloaded as Markdown or a zip including the research and the agent actions.

Sessions no longer expire unexpectedly. Older sign-in cookies are converged onto the current token chain, so people are no longer signed out part-way through the day without a deployment having happened.

Reliability. Long-running work keeps its answer instead of losing it to a database timeout; a single malformed document can no longer block document conversion for everyone; document and spreadsheet generation from the Workbench works again; and removing a role from an administrator no longer returns a server error.

Interactive canvases in chat. The assistant and skills can now open a live, interactive panel — a chart, an editor, a small app — in a split view beside the conversation. Keep chatting to steer what the canvas shows; canvas versions are saved with the thread and survive a reload. Admins decide per space whether Canvas is available.

Share a conversation. Share any chat read-only with your whole space or with specific members. Recipients see the conversation exactly as it ran, without being able to change it.

Aimable Dictate: talk instead of type. Push-to-talk voice input in the composer — hold the microphone button (or hold Space) and speak; your words land in the composer as text. Available once an admin enables Dictate for the organization.

Transcription dictionary. Teach the transcriber your names and jargon: a custom vocabulary and correction dictionary is applied to dictation and meeting transcription. Spot a misheard word? Correct it inline and add it to the dictionary from right there.

Sources you can actually open. Citation rendering has been overhauled: knowledge-search sources now link to the underlying documents, and citations behave consistently across languages and across follow-up turns.

A tidier composer. The composer toolbar is consolidated into a + menu (attachments, skills, and other inputs) and a settings menu, so the chat box stays clean as capabilities grow.

The assistant shows what it's doing — continuously. Long turns no longer go quiet: activity indicators keep streaming while the assistant reads files, searches, or thinks, and errors that end a turn are shown as a proper error state instead of a note buried in the text.

Images in chat history are clickable. Thumbnails on earlier messages open in a full-size viewer.

New chats remember your knowledge scope. A new conversation starts with the collection selection you last used in that space, instead of resetting to nothing.

Word documents keep their look through editing and translation. The document editor now handles bilingual .docx work while preserving the source document's fonts and numbering.

Skills can save their deliverables into a collection. A skill's output can be persisted straight into a knowledge collection, so results are findable later instead of living only in the chat thread.

SharePoint connector. Connect a SharePoint document library and keep it in sync as a collection. The connection now stays alive without re-consenting (refresh tokens), and each organization can use its own Microsoft app registration.

Bot-protected websites can now be indexed. Website collections fall back to a real browser for sites behind aggressive bot protection, so pages that previously came back empty now index correctly.

Cost dashboards, sharpened. Model charts on the tenant and space dashboards now show real dollar amounts, and a new Cost by API Key chart breaks spend down by key and provider. Trace counts are labeled "Interactions" for clarity.

Per-space Canvas toggle. Decide per space whether the interactive Canvas tool is available, from the space governance settings.

Dictate is opt-in per organization. A new feature toggle in Console enables Aimable Dictate for your tenant. Action needed if your users should get voice input.

Organization transcription dictionary. Manage the org-wide custom vocabulary and corrections used by dictation and meeting transcription from a dedicated Console page.

Skill-execution caps per tenant. Configure how long and how many iterations a skill run may take (wall-clock and step limits) from tenant settings.

Tenant Admins are now visible. The principals list shows a Tenant Admin badge, just like Global Admin, so it's clear at a glance who administers the organization.

Usage export knows who spent what. usage_by_model is enriched with the provider and the API key behind each interaction, enabling per-key cost attribution downstream.

No more silently unpriced models. The model catalog warns when a model has no pricing configured (instead of billing €0), and pricing was added for scaleway/qwen3.6-35b-a3b.

File content endpoint accepts thread_id. The document content endpoint can resolve a thread's collection directly from the thread id, so clients no longer need to look up the collection first.

Office add-in: sign in with Microsoft — or any configured provider. The Word & Excel add-in now discovers your server's identity providers instead of assuming Google, and the Microsoft sign-in popup issue is fixed.

Office add-in: switch organizations. Users who belong to multiple tenants can switch tenant from the add-in settings panel.

Office add-in: clearer status and settings. The busy indicator now stays visible for as long as the assistant is working, settings show which server you're connected to, the header is decluttered (follow-mode and theme moved into settings), and individual chats can be deleted from the session list.

Aimable Capture (in development): a redesigned meetings experience. The meetings list and detail pages have been overhauled, chatting with a meeting is now the primary action, and a context chip in the chat shows exactly which meeting the conversation is focused on.

Aimable Capture (in development): better summaries and transcripts. Summaries of long meetings no longer come out empty, meeting content is retrievable from regular chat, a meeting can be re-transcribed in the background, and transcripts show participant names instead of anonymous ids. The transcription dictionary is applied to meeting transcription too.

Aimable Capture (in development): concurrent bot limit. How many meeting bots may run at once is now configurable per tenant, with a Bot limits card in the Console meeting settings.

JSON files upload correctly. .json uploads to collections are now indexed (previously accepted but silently dropped), unsupported file types are rejected with a clear 415 error, and the upload dialog lists exactly the formats the pipeline accepts.

Rich table content survives import. Tables with formatted cells no longer lose their content when documents are indexed into knowledge.

Skills stay grounded in your real documents. Skill runs now stage the original files of your selected collections and the thread's earlier outputs into their workspace, and are explicitly forbidden from reconstructing a source file they don't actually have — no more plausible-but-invented content.

Space members can run skills. Skill execution now honors space-level permissions instead of being blocked by a tenant-level gate.

Chat titles match your language. Auto-generated conversation titles are pinned to the language of the message itself.

Composer '/' cleanup. Dismissing the skills menu without picking a skill no longer leaves a stray "/" in the composer.

Document downloads fixed. Generated documents no longer intermittently 404 on download, and document tools are more tolerant of the formats models hand them.

Indexing no longer destabilizes the platform. A bug that drove full re-indexing loops and out-of-memory restarts during large document syncs has been fixed.

Clear error on invalid API-key minting. Minting an API key for a human principal returns a 422 with an explanation instead of a bare 500.

Connector OAuth on multi-tenant hosts. OAuth callbacks are now built from the tenant's own host, so connector authorization works correctly on multi-tenant deployments.

Per-tenant web-search keys. Web search no longer runs on a shared global key — the new Integrations tab in Console Settings holds each tenant's own Exa key, and a tenant without its own key has web search disabled. Action needed for every tenant that should keep web search.

PII mappings moved from X-Pii-* headers to the request body (metadata.pii_mapping), eliminating 431 Request Header Fields Too Large on long conversations; the header path is kept as a fallback during rollout.

Skills can now ask you questions mid-run. When a skill needs more input, a question panel appears above the composer (one question at a time — no more stacked walls of questions). Answer to continue, confirm free-text answers with Enter, or decline a question to cancel the run. The skill card's live step list keeps updating through these pauses.

Skills no longer re-ask for information you already gave in the chat. The conversation context is passed into chat-invoked skills, so data mentioned earlier (or fetched by another tool) is picked up automatically.

Type / in the composer to discover and start skills — a slash-command menu lists the space's skills by display name; picking one inserts a skill chip.

Skill results are now reliably visible: generated files appear as download chips even when the skill returns no chat text, progress is surfaced while the skill runs, and skills that dispatch sub-skills show each sub-agent's progress live in the chat.

Skill bundles work on multi-tenant: tenants can import skill bundles (SKILL.md format, Anthropic-compatible, including packaged .skill files) and attach them to spaces — with code execution disabled on multi-tenant as a security boundary. See Building and Deploying Skills.

Skills now run on the space's active model instead of silently falling back to the space default.

Pick which collections are in scope per conversation: a selector next to the composer lets you toggle the space's collections on/off for the current thread (e.g. search one legal corpus only). Space admins enable this per space.

A website can now be indexed as a collection — point a collection at a site and its pages become searchable knowledge.

Knowledge search got cheaper and cleaner: query-aware compression and chunk filtering cut the tokens sent to the answering model, and duplicate chunks no longer appear in results.

Aimable now knows today's date — the current date is in the system prompt, so time-sensitive questions ("this quarter", "last month") resolve correctly.

Spreadsheet content is no longer sent to the PII detection server. Excel/CSV cell data stays out of the contextual scanner across preview, chat upload and knowledge flows.

A more thorough optional PII check per space, with a clearer in-chat experience while it runs, and your manual corrections to detected PII are now applied exactly as you made them.

Organisation names (OII) now follow the space's pseudonymisation setting in skills too, with a per-skill manifest override — previously skills always pseudonymised organisations regardless of the space setting.

Switching tenants now takes you to that tenant's own URL, so sessions and bookmarks land on the right subdomain, and deleting the conversation you're viewing returns you to a fresh new-conversation page.

Settings view redesigned, now including the tenant's URL.

Model management hardened: deactivating a catalog model is now validated against space model policies instead of silently breaking chat in those spaces; deleting and re-adding a tenant model with the same alias works again; GLM 5.2 and DeepSeek added to the catalog.

Cost & observability accuracy: the observability dashboard no longer caps traces at 100, under-reports cost, or hides space-less traces; cached-token costs are shown correctly, including Tensorix cache reads that were billed at $0; Tensorix models no longer display with a misleading openai/ prefix.

Platform admins can now open any space collection and run skills in any space without being an explicit space member.

X-Prompt-Cache header on /v1/chat/completions enables Anthropic-style prefix caching on the stateless completions endpoint; model pricing is now exposed in the model-policy API.

Admin usage-export endpoint returns tokens (in/cached/out), cost and model per tenant/space per period, including tenant_name, for billing tooling.

knowledge_search execute endpoint now honours collection_id — previously it silently searched all space collections, bleeding results across collections.

Download links survive platform redeploys — skill and generated-document artifacts moved off ephemeral storage (no more 410 Gone).

Chat surfaces a clear inline error when a prompt exceeds the model's context window, instead of silently returning nothing; a malformed tool call from the model now retries instead of aborting the whole chat turn.

Citation source downloads were broken on /api-proxy deployments (double path prefix); skill artifact download chips now download via authenticated fetch, and skill cards no longer poll endlessly on dispatch-level failures.

Free-text clarifying questions no longer crash the chat, and sequential questions no longer freeze after the first answer.

Knowledge indexing robustness: knowledge_search no longer fails with gRPC >100MB errors caused by quadratic PII-span duplication; PDF table extraction no longer drops spaces inside table cells, fixing corrupted entity names in indexed documents; markdown indexing failures for collections on the multi-tenant server fixed.

May API pentest resolved. All findings from the May external pentest (medium/low, nothing high) are fixed.