AI Usage Plans, a Rebuilt Spark Editor & Artifact PDF Export

Copera's pricing becomes three simple layers — a platform that is free forever for unlimited members, one AI usage plan per workspace with a weekly allowance that resets, and add-ons you buy one at a time. Alongside it: a rebuilt Spark editor where the camera is finally its own layer you can move after recording, PDF export for Artifacts, and an Ask Copera that remembers what you tell it, branches from any message, and lets you undo an action it already committed.

New Features

AI Usage Plans — A Free Platform and a Weekly Allowance That Resets

The old Plus/Max Block model is gone. In its place, three layers you can hold in your head (#3528, #3494):

• The platform is free forever, for unlimited members — boards, docs, drive, chat, meetings, inbox and the rest. Adding people to your workspace never changes your bill, and there is no seat pool to run out of.
• One AI plan for the whole workspace — a ladder of twelve plans, each named for how much AI it gives you compared to the free tier (1x through 600x). You pick one per workspace, not one per person.
• A weekly allowance that resets — your plan grants an amount of AI use every week, and it starts fresh at the next reset instead of carrying over. A week that gets away from you is one week, not one month.
• A slower lane instead of a hard stop — when a paid workspace spends its week, AI keeps working through a slower queue rather than shutting off. Free workspaces stop until the reset.
• Extra credit when you need it — buy more at any point once the week is spent, and go straight back to full speed.
• Add-ons a la carte — extra inboxes, extra storage and live classroom capacity are each bought on their own, on any plan including the free one, at a flat monthly rate.
• Better models on bigger plans — the heaviest and most capable models unlock as you move up the ladder, and the model picker tells you how long your week lasts rather than quoting rates.
• See where you stand — usage bars in the app show how much of the week your workspace and you personally have used, with the days left until reset (#3531, #3530).
• A rewritten pricing page — the new plans page walks you through a week of real work on each plan, plays out what happens when the ceiling arrives, and drops the per-unit dollar figures that people kept reading as per-seat pricing.

Sparks — A Rebuilt Editor, a Camera You Can Move, and Local Rendering

The biggest change Sparks has had (#3524). Recording and editing are now separate concerns, so decisions you used to make before pressing record are decisions you make afterwards:

• Edit on the clock you are watching — split, ripple-delete, trim, reorder, change speed, adjust per-clip audio, crop, and add zoom regions, all on the finished timeline rather than on raw source. A cut collapses the gap instead of leaving a hole.
• The camera is its own layer — your webcam is no longer painted into the recording. That means the bubble can be moved, resized and restyled after the fact, and sharing a window that is not Copera no longer loses the camera entirely.
• What you preview is what you export — the preview and the final render now compute the same frame, so zoom, crop, camera placement, padding, rounded corners and speed all look on screen exactly as they will in the file. Zoom eases in and out like a camera move rather than snapping.
• A cursor you can restyle — on the desktop app the pointer and its click rings are recorded separately from the picture, so click highlights stay editable instead of being burned in. Only position, timing and whether a click happened are captured — no keystrokes, no window titles, no app names.
• Render on your own machine — the desktop app can produce the final file locally and publish it back to the Spark, with the server checking the result is the file it was promised before anything replaces your existing cut.
• A contextual inspector and keyboard shortcuts — controls appear only for what you have selected, speed is offered as the exact set of rates the renderer supports, and shortcuts are ignored while you are typing so renaming a Spark can no longer split a clip.
• Autosave that does not lose work — edits save against an explicit version, an edit made while a save is in flight is never silently dropped, and a genuine conflict asks you to reload and reapply instead of sending you to retry into the same wall.
• Screen plus camera, where it works properly — the combination is offered on the desktop app always and on browsers that can record both streams at once; elsewhere it falls back cleanly rather than sitting selected and disabled.

Artifacts — Export to PDF

Any Artifact version can now become a PDF (#3550, #3552):

• One click to export — export from the studio and get a file back, with a receipt that stays on screen instead of vanishing into a four-second toast.
• Portrait, landscape or automatic — a handle beside the export action lets you force page orientation. Presentations keep their own aspect ratio and are never forced.
• Presentations render every slide — a deck exports one page per slide rather than a single page with the rest hidden.
• Dashboards and apps print at their real size — dashboards, tools, visualizations and web apps print at their actual screen layout with inner scroll areas expanded, instead of being squeezed into a page and clipped to one screenful.
• Open or download, on desktop too — Download saves the file on both web and the desktop app, and Open now works in the desktop app instead of failing every time.
• It stays available — a rendered PDF remains reachable from the studio's menu for its full seven-day window, across reloads, and is always the PDF for the version you are actually looking at.

Artifacts — Full Screen and a True Split View

The workspace had exactly one shape: chat as a side panel. Now it has three (#3539):

• Full screen — hand the Artifact the entire viewport with no chat and no app chrome, and press Esc to come back.
• Split view — an even 50/50 split with the Artifact on the left and the chat on the right, automatically collapsing the left navigation to buy back the width. Leaving restores only what split view changed; a navigation you had already collapsed yourself stays collapsed.
• No more reloading — toggling in and out of full screen used to reload the Artifact and lose your scroll position and anything mid-animation. It no longer remounts anything.
• Neither mode strands you — both step back to the panel layout the moment the window is too small for them.

Ask Copera — Undo an Action It Already Committed

After Copera commits a reviewed action, the card now offers Undo and the workspace goes back to how it was (#3544):

• What can be undone — row updates, created rows (single and bulk), created columns and label-only column edits.
• What says so honestly — deletes, created tables, and column edits that changed options tell you plainly that they cannot be reversed, rather than pretending.
• It refuses rather than surprising you — if the thing changed since Copera wrote it, the undo stops and says the target has moved on instead of overwriting a colleague's edit.

Ask Copera — It Remembers What You Tell It

The memory in AI settings was real, but only you ever wrote to it. Now a finished conversation can capture a durable fact about you — a standing preference, your role, how you like things done (#3538):

• Only durable facts — conversation summaries, one-off task details, facts about other people and anything you merely pasted are all deliberately left out.
• Never anything secret — anything that looks like a password, key or token is refused outright, in every product language, whatever the rest of the message says.
• Everything stays yours — captured lines are marked as written by Copera, sit in the same free-text field you already edit, and can be changed or deleted one at a time. Editing an AI-written line makes it yours.
• It stops rather than forgetting for you — at the twelve-entry cap capture simply stops. Deciding what to forget is your call.
• Edits and captures no longer collide — saving the memory field while a capture lands mid-edit now merges the two and asks you to review, instead of one silently overwriting the other.

Ask Copera — Branch From Any Message

Fork a conversation from any point, edit a message and re-run it, regenerate a reply, and move between the versions with a small picker on both your message and Copera's reply — and all of it survives a page refresh (#3541).

Ask Copera — It Tells You When a Run Needs You

• When a turn finishes without you — a long turn that completes after you have walked away now raises a notification and a mobile push. You are not notified for a reply you watched land, and a turn that chains several steps is one piece of news, not four (#3526).
• When a run is waiting for your approval — the one state that genuinely needs you back was the one state that stayed silent. It now sends a notification and a push, and the bell colours the outcome honestly: green for completed, yellow for waiting on you, red for failed (#3535).
• One switch for all three — a single preference covers finished, failed and waiting-for-approval, and says exactly that.

Ask Copera — Ask for a Dashboard, Get a Dashboard

Asking for a dashboard used to quietly produce an Artifact instead, because building a real Copera dashboard was not something Copera could do. It is now a proper reviewed action: it creates the dashboard under your own permissions and can fill in the widgets from your description (#3543).

AI Workforce — Weekly Schedules and the Chat's Full Toolkit

• Agents can run weekly — "every Monday at 9" is now a real schedule instead of an approximated interval, so a weekly report is one job rather than seven. The schedule keeps its wall-clock time across daylight-saving changes (#3527).
• Agents can build Artifacts — every agent surface can now author an Artifact, not just scheduled jobs, and the artifact is always bound to the person who hired the agent (#3542).
• The same tools the chat has — automations, agent jobs, the command palette, inbox, to-do planning and the bulk board tools are all available to agents now, each behind its own capability so an existing hire sees no change until you grant it (#3545).
• They say something when they stop — an agent that hits its step limit now wraps up with a summary instead of ending in silence, and its step list reads in your language.
• Approval means approval — an agent set to ask before writing unattended used to go ahead anyway on anything that was not destructive. Every write now waits for you.
• Failures come with a reason — a scheduled or automation-triggered run that fails now produces a receipt explaining what went wrong, and the reason reaches you in the notification.

Inbox — Start WhatsApp Conversations, and Threads That Behave

• Reach out first — start a WhatsApp conversation from the inbox by typing a number, on any connected WhatsApp channel. The thread is assigned to you from the start, and replies land in the same conversation (#3551).
• Closed means closed — a closed conversation now stays closed and the next message opens a fresh thread, instead of reopening a ticket you had already finished (#3553).
• One person, one conversation — replies that arrive under a different identifier for the same number are now merged into the conversation you already have, rather than opening a second one (#3553, #3556).
• Media and reactions — photos, videos and documents now open reliably, reactions show up as reactions rather than empty bubbles, and empty bubbles left behind by earlier deliveries have been cleaned up (#3553).
• Send with your name — a new channel setting controls whether your name is sent alongside your messages.

Developer — Inbox Channels and Conversations in the Public API

• New endpoints — list your channels, conversations and messages, reply to a conversation, and start or send a WhatsApp message to a number, all behind a dedicated Inbox permission you grant explicitly (#3551).
• Also in the CLI — the same operations are available as commands.
• Calendar writes notify by default — creating, updating or deleting an event through the API now emails the participants unless you explicitly ask it not to, which is what bulk imports and syncs should be doing.

Bug Fixes

• Big tables stopped loading past about 32,000 rows — the ordering value overflowed and every row past the limit produced an error. Tables of any size now load normally (#3554).
• Artifacts that rendered as raw text with no styling — an Artifact could be published with its internal packaging as the page itself, so it appeared unstyled with stray characters everywhere. Two people read it as Copera having become a worse designer; it was never the design (#3534).
• A review card that failed took the whole answer with it — a turn where every tool had already succeeded could still die with "Invalid request" because a single approval card could not be parked. The card is now dropped and Copera answers in text instead (#3533).
• A busy model provider killed the turn permanently — a rate-limited provider, the most temporary failure there is, was being treated as a permanent dead end. It is retried properly now (#3533).
• Turns that ran out of time said nothing — a run that hit its twenty-minute limit died at the boundary and showed "Internal error" with nothing to show for the work. It now reserves time at the end to answer with what it built, and a timeout is reported as a time limit rather than an internal fault (#3547).
• Error messages arrived in untranslated English — every server-side error the app showed was falling through to the raw internal English message, even though translated copy for it existed in all six languages. Coded errors now read in your language again (#3547).
• "This version cannot be published" with no reason — the share dialog answered four different situations with one sentence. It now names the case: still building, failed for good, or held back by the design review — and shows it inline instead of "The publication review could not be loaded" (#3547).
• The chat inside the Artifact workspace could take over the screen — a one-frame glitch in the embedded chat escaped to the application error screen, replacing everything with "Refresh / contact support" even though the run had completed correctly. It now recovers on its own (#3529).
• Automations reading a description field — a rule using a row's Description column received internal storage data instead of the text. It now gets the actual written content (#3546).
• WhatsApp channels stuck on "Pending" — a channel could never finish connecting because some provider events carried no timestamp, and reconnecting could show a stale phone number. Both fixed, along with QR codes that refused to refresh and a pairing that two people could fight over (#3548, #3551).
• WhatsApp messages silently not arriving — incoming deliveries were being rejected, so nothing showed up at all. Channels also now tell you plainly when a number is already registered elsewhere instead of failing vaguely (#3549).
• Ask Copera losing the board it just made for you — a conversation whose only context was a board Copera had just created refused every follow-up edit and told you to do it in the app. It now remembers what it built, and the board you have open counts as the board you mean (#3559).
• Ask Copera writing to the wrong table — a follow-up could land on a neighbouring table of the same board that you never named. It now refuses unless you named it (#3559).
• AI-created rows waiting on the next sweep — formulas, linked records, search and deadline rules did not update on a row Copera created until a background pass caught up. They run immediately now (#3559).
• Link previews with no image — pages on the website could be shared with no preview image at all; there is now always one.

Improvements

• Usage bars speak in percentages — the weekly bars no longer show dollar figures, the heading sits above the bar with the percentage and the caption below it, and the extra-credit offer appears only once the week is actually spent (#3531, #3530).
• Legacy subscriptions show their own price — workspaces on an older subscription no longer see the new ladder's price quoted as though it were what they pay (#3530).
• Failures are named instead of guessed at — Copera's own errors now carry their own explanation through to you, the receipt says which stage actually failed rather than always blaming the model, and the thread shows a Details line in all six languages (#3540).
• If a card could not be created, Copera says so — instead of promising an approval card that never appears, it now tells you plainly that the action was not applied and nothing is waiting on you (#3540).
• A run that continues in the background is never lost — a follow-up step that was accepted but never started is now picked up and dispatched rather than being quietly abandoned (#3543).