What has changed?

Messages, replies in threads, mentions and attachments, plus user mapping and organization settings.

The coworker reads the channel and acts in it on its own, on the same terms as in Slack.

The bot reads pasted images and files from the channel, including ones kept in SharePoint.

A Teams user is linked to their appto account by email address, with no manual mapping.

Its own settings, access and conversation history.

GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Claude Opus 5 and Grok 4.6. Luna is also available in the assistant builder.

Added: Ahrefs (domain analysis and keyword metrics), Exa (deeper research with the content of the pages found) and LinkedIn (searching public posts).

A task in a Project can be assigned to a specific project member.

Preset values to choose from or your own — from a minute to a day before the event.

The statistics no longer mix the two messengers.

The bot introduces itself on the first message, replies in the same thread and says plainly what to do when the organization is not connected yet.

Instead of cutting the reply off at “I will try again”, it more often tries once more and sees the thing through.

A taken name, a missing required integration or missing permissions — the message says what to fix.

The icon next to a tool now explains exactly what it means.

It used to revert to Wise.

The working animations stop at once and the partial reply is saved correctly.

This used to end in a server error.

The coworker remembers who wrote which message in the conversation.

You type a command, the thread appears at once, the agent works in the background and the input frees up immediately — you can delegate the next one right away.

Finished and unreviewed items land in For you. Plus Pinned and a Snoozed shelf: in an hour, this evening, tomorrow at 9:00 or on Monday.

j/k to navigate, Enter opens, e marks as done (with undo), s snoozes, f full screen, 1–9 jumps to a row. Holding Ctrl shows the hints.

Apollo (lead search and enrichment), People Data Labs (data on people and companies), X, Reddit, Facebook and Google Reviews. They run on our side — no sign-in and no account with the provider.

A single call costs 300, 500, 800 or 1500 credits, depending on the tool.

In advanced settings, on a paid plan. The user sees them grouped by provider.

The admin sets whether an integration is on, for whom — the whole organization or selected teams — and whether it is required.

One shared company account per integration, or many personal ones with aliases.

You can see who has not connected what — with filters and a search box.

Someone with a verified address on the company domain joins in one click, with no invitation and no admin — as long as the plan has a free seat.

The same rules in each of them, instead of four separate paths.

Without going back to the catalog and confirming a second time.

A search box, categories and cards instead of long lists.

Reasoning models used to take one tool per round and every action cost a full lap.

The model gets a preview and reads the whole thing only when it needs it.

Replies in a thread stopped getting lost, and the bot reacts to thanks and jokes with an emoji instead of staying silent.

An artifact clicked in the drawer opens straight into full screen.

A sleeping laptop or a dropped connection no longer freezes the view — it comes back within tens of seconds.

The panel also stopped opening by itself.

Even when the second attempt gets different data than the first.

It no longer takes over a discussion someone else opened.

Before content reaches the model, names, email addresses and national ID numbers turn into placeholder tokens — [PERSON_1], [EMAIL_1], [PESEL_1]. The appto chat model never sees data in the clear where the organization’s policy covers it.

A switch in advanced settings, on the Privacy tab. Anyone who does not touch it sees no change at all.

The file shows its processing state, and masking is reapplied on edits and on writes made by the agent.

The gateway either processes the file or stops the upload — there is no third path.

The token-to-value pairs sit encrypted in the Privacy Vault for 30 days and stay out of the model’s reach.

Files never leave the EU. Stated plainly on the Privacy tab.

Charged once, after successful processing. Reusing the same file, or having the mode off, costs nothing.

You see how much data the gateway recognized, and of what type.

The bot scans channels and decides for itself when it is needed. It used to answer only when spoken to.

Channel by channel or all at once, automatically covering new ones. The mode covers channels only.

A channel with no new messages costs nothing — the bot does not even ask the model.

Outgoing, Balanced or Reserved — the default, reproducing the behaviour from the pilot. The setting moves confidence thresholds and per-scan limits together.

It goes the same route as a reply to a direct mention.

A run of failed charges switches autonomous mode off automatically.

With no manual adding, including installations older than the current Slack notification mechanism.

Attio is a CRM. Apify runs ready-made data-gathering robots — a new category in the catalog.

A context of over a million tokens, text and image. Only US-hosted providers serve it. Priced like the Sonnets: 10 credits per 1000 input tokens, 50 per 1000 output. Also available as a Wise model — Sonnet 5 remains the default.

For billing admins in free organizations: the ring goes from green through amber to red and leads straight to billing.

It writes normally in the text, and the displayed reply substitutes the real values wherever they may be shown.

Attio writes to the customer’s CRM and running an Apify actor costs money — both fall under the policy that requires approval for writes.

Instead of a Pro toggle — a model list that opens next to the message box.

Useful where the answer should be repeatable rather than depend on whatever someone happened to pick.

It creates and reads them, returns binary files too, and a separate card shows how the run went.

Less waiting for the first word and for the reply to close.

Its price in credits changes.

A new module: the routine starts by itself when whatever you set up happens.

A prepared trigger starts switched off — you turn it on deliberately.

The place where you assemble and test a routine before it starts running on its own.

A new project is created from a ready template: with a space, a walkthrough and a choice of assistant.

Inside a project you choose from that project’s agents.

The strongest reasoning in the Sonnet family. Sonnet 4.6 remains available.

A large context window, strong on agent tasks and on code. Only providers from our approved list serve it.

A manually chosen setting is not overwritten.

A new module in account settings: who did what and when — agents, skills, contexts, billing, members, the Slack bot and security, each area with its own detail panel.

Multi-select plus view tabs.

Instead of raw technical names — a sentence that says what happened.

The log records the event, not the content.

Indicators, charts and rankings, plus detail windows for every metric.

Behind the name sits a new engine and a rebuilt module.

With editing and cascading deletion. A brief is required.

Without sending the task anywhere else. Code mode is the default.

The bot reads but changes nothing.

Along with new defaults for new customers.

Instead of finding out from a failed task.

Chat, sign-in, the greeting and the app switcher — alongside account settings, which got a Polish version first.

A broken stream shows a message, not an empty page.

Text from outside — from a page, a file, a message — is data to the model, not an instruction.

Installation in the workspace, replies in threads, recognizing who is writing, and knowledge set separately for each channel.

Anyone without an account yet learns from the bot what to do.

A lighter variant on a fixed model. Wise stays the default, on Sonnet.

The admin picks which models may be used.

A new conversation starts with the right assistant straight away.

Such a conversation is shared within the organization.

An interrupted connection can be retried without starting over.

VAT and the tax number are calculated automatically.

In the chat and in account settings. The choice holds across subdomains.

Instead of starting the reply over.

This used to end in an error halfway through the reply.

The session is kept between calls instead of being opened every time.

A shared space with threads where several people work at once.

PDF, DOCX, XLSX, Markdown and images upload into the project; the content is read and image descriptions are billed as normal usage.

Everyone sets listening for themselves; otherwise the agent speaks up when mentioned in a message.

The file is yours — it is appended to the prompt and stays exactly as you save it.

This used to require organization management permissions.

You can upload it again straight away.

While condensing, sending is paused, and a failed condensation is shown plainly instead of failing silently.

Instead of stopping halfway.

Instead of failing without explanation.

A Markdown instruction the assistant reaches for when it is needed. Creating, editing and reference files with a preview.

Chosen in the builder.

Once an account is connected the assistant gains a skill for that tool. The first one covers Gmail.

From this release the whole conversation, streaming and tools run on the new environment.

A scanned PDF of up to ten pages is read as images.

Warsaw time, supplied up front — one call fewer in every conversation.

Saving the conversation stopped blocking what you see on screen.

Up to 300,000 characters per file and 600,000 per message.

The interruption happens in a fraction of a second.

The order of text and tool calls is preserved live too.

Upload, processing and download all stay within the organization’s boundary.