Conol 0.53.0 — Your workspaces, in your order

Most of this release is about things being where you left them: a workspace list that stops rearranging itself, an account badge that shows the account you're actually in, a note title a viewer can't type into. The last section isn't like the others — it's the credit and premium-tool changes that came out of the bot-farming waves we wrote about this week.

Put your workspaces in the order you want

Why. The request that started this was simple: workspaces could only be listed in the order you joined them, and there was no way to change it. Reproducing it turned up something worse underneath. The list had no ordering rule at all — not join order, not alphabetical, nothing. It looked like join order most of the time, which is the whole problem, because the moment any membership record was rewritten (someone's role changed, or an ordinary background touch of the "last updated" field) that workspace moved, and your list silently reshuffled with no action from you. Two loads of the same page could hand back two different orders.

If you live in one workspace you'd never notice. If you have six and navigate by muscle memory, the list moving under you is a small, constant tax.

How to use it. Open Settings → Workspaces. Each row now has Move up and Move down arrows; click your way to the order you want and it's saved as you go. The order applies everywhere your workspaces are listed, including the sidebar, and it survives a reload.

Four details worth knowing:

• The order is yours alone. A workspace is shared, so one member rearranging their list must not move it for everyone else. Your order is stored against your account, not against the workspace.
• The default is now stable. Without a saved order, workspaces are listed oldest membership first, with a deterministic tiebreak so two workspaces joined at the same instant can't trade places between requests.
• Nothing can fall off the list. A workspace created or shared with you after you set your order is appended at the end rather than dropped. A workspace you leave is skipped rather than breaking the order.
• A failed save says so. If the reorder doesn't reach the server, the list snaps back to what it was and shows an error, instead of leaving you looking at an order that isn't real.

GLM 5.3 joins the model picker

Why. Z.ai released GLM 5.3 on 2026-08-18. It's the same base model as GLM 5.2 with a much heavier post-training pass, and the gains are concentrated in coding rather than spread thinly. It carries the same one-million-token context window as GLM 5.2.

The practical reason to care is the price. GLM 5.3 blends to roughly half our baseline model's per-token rate, which keeps it under the multiplier at which models become premium — so it's available on a free account, and it's a genuinely strong coding model to have there.

How to use it. Open the model menu in any chat and pick GLM 5.3 under Z.ai. Two things behave differently to most models:

• It always reasons before it answers. Its endpoint rejects a request that disables thinking, so there is no non-thinking mode to fall back to — pick the Effort level that suits the task, or leave it on Default and let the model choose.
• It's text only. No images or video in, so a screenshot or a photo needs a different model. Notes, code, long documents and large sets of text are what it's for.

GLM 5.2 stays exactly where it was for anyone who prefers it.

Agents can reach a workspace you create mid-conversation

Why. When a conversation started, your agent was handed a list of the workspaces it could reach — a snapshot, taken once, at that moment. Everything after that was invisible to it. Ask your agent to set up a new workspace and then put notes in it, and it would create the workspace and then insist it had no access to it. The only way through was to start a new conversation, which meant losing the thread you were in the middle of.

The frozen list was never a deliberately narrower permission — it was meant to be "every workspace this person is a member of", captured once for convenience. Freezing it bought nothing and cost three kinds of staleness: a new workspace was unreachable, a deleted one lingered, and a membership someone had revoked lingered too.

How to use it. Nothing to turn on, and nothing to restart. Ask your agent to create a workspace and then write to it in the same breath; ask it to list your workspaces and the one you just made yourself in another tab is there. Membership is now checked at the moment of each request rather than remembered from the start of the chat.

This grants no new access. It's the same set of workspaces that was always intended, evaluated now instead of an hour ago — and it's strictly tighter in the directions that matter: if your access to a workspace is removed mid-conversation, your agent loses it immediately rather than at the end of the session.

Running chats sit with your pinned ones

Why. The sidebar filed running and scheduled sessions together under one heading, below your pinned ones. Those are different things to a reader: a running session wants your attention now, a scheduled one is a standing arrangement that will fire on its own. Mixing them meant the thing you most likely opened the sidebar for — the chat that's working right now — could sit halfway down the list.

How to use it. Nothing to turn on. Anything currently running now appears at the very top of the list, in the same section as your pinned sessions — running first, then pinned in the order you arranged them, so the shelf never reshuffles under you. Scheduled runs get their own Scheduled section just below. A pinned session that happens to be running shows once, with its running indicator, right where you pinned it.

Subscription renewals grant their credits again

Why. This one is ours and it's bad. Stripe changed where an invoice records which subscription it belongs to. Our renewal handler was reading the old location, found nothing there, and therefore couldn't tell which subscriber the renewal was for — so it granted nothing and moved on without erroring. Payment succeeded, the subscription stayed active, and the credits never arrived. Every renewal since launch granted zero credits.

It was invisible from the inside because nothing failed: no error, no retry, just a missing ledger entry that nobody was watching for. We found it while auditing the billing system for something else entirely.

How to use it. Nothing to do. Renewals now resolve the subscription correctly and grant the credits your plan includes. Every affected subscriber has already been credited by hand. If you renewed at some point and felt your balance didn't move, you were right — and if you think you were missed in the backfill, write to us and we'll check the ledger.

Viewers can no longer edit a note title or emoji

Why. If you were a viewer in a shared workspace, you could click into a note's title and type. It accepted your keystrokes, appeared to save, and then reverted to the old title the next time the page loaded. The same went for the note's emoji.

The server was never actually confused — it refused the change every time, exactly as it should. The failure was that the refusal went nowhere: the title box sent its update and never looked at the answer. So the product showed you an editable field, took your input, discarded it, and said nothing. Read-only had already been applied to the note body itself; the title and the emoji picker were simply left behind.

How to use it. With viewer access, the title is now read-only and the emoji picker is disabled — the affordance matches the permission. Owners and editors are unaffected. This also covers the brief window right after opening a note, before the collaboration connection has confirmed your role: no title save is sent while the answer to "may I edit this?" is still outstanding.

The account you are signed in as is the one on top

Why. With more than one account signed in, the sidebar shows their pictures stacked and overlapping, current account first. The list order was right — but overlapping images with no explicit stacking order are painted in the order they appear, so the last account was drawn over the others and the current one sat at the back. Because the current account is always first in that list, it was always the one underneath — every time, for everyone.

For a feature whose entire purpose is knowing which account you're in before you type something into the wrong one, the badge pointing at the wrong account is worse than no badge.

How to use it. Nothing to do. The account this tab is pinned to is now drawn in front of the stack, and stays left-most, so both readings of the badge agree.

Changes to free credits and premium tools

Three changes to the free tier ship in this release. All three come out of the credit-farming operation described in The week the bots came — that post has the numbers and the reasoning. Here is what actually changes for you.

The daily free allowance drops from 300 credits to 100. It still refills every day, forever, and 100 credits is still several real conversations a day on capable models. It is also, plainly, a reduction: if you were using the full 300, you now have a third of that.

Referral invites unlock one week after you create your account. The program is otherwise unchanged — both sides still get 500 credits when an invite is accepted. New accounts can't send invite links yet; the invite panel shows the date yours open. Share links from an account younger than a week still work as share links, they just don't carry a referral bonus, and the sign-up page no longer promises credits it wouldn't grant. Bonus-chaining between fresh accounts was the mechanism behind roughly an account a minute, and a seven-day wait is the smallest rule that breaks it.

The Code tool and sandbox access to model providers now require a paid plan — a subscription or a credit purchase. This is the change we're least happy about. The Code tool hands your agent a full cloud computer with frontier coding agents inside it, reaching the model providers through our gateway; that same shape is, structurally, close to an unmetered API key, and it was the path most of the farmed spend flowed through. The previous exception for people supplying their own coding credentials is gone too — it was being used as a way in rather than as intended. Enforcement now happens on the server where the requests are actually made, not just in the interface, and it applies the same way in Telegram, Slack, WeChat, Discord and WhatsApp as it does on the web.

Everything else in the free tier is untouched: daily chat with capable models — GLM 5.3 among them — notes and the editor, agent memory, web search, timers, channels, and the whole collaboration side of the product.

Fixes

• Losing workspace access takes effect immediately — a membership revoked mid-conversation used to keep working for the agent until the conversation ended.
• Share pages stop promising credits they can't grant — a share link from an account too new to earn referral credits no longer advertises "both get credits" on the shared note or the sign-up banner.
• Usage that was going uncounted now counts — long-lived prompt-cache writes, auto-generated chat titles, page summaries and connector calls were being recorded at a cost of zero. They're now billed at what they actually cost, which is a small increase in what a heavy session records.
• Banned accounts stop running work — a banned account's live sessions are cancelled and its scheduled timers removed, instead of continuing to run and spend in the background.
• The Terms of Service describe the referral rules that now apply, including the one-week unlock.