Changelog

A signed-in account can now make 40 designed documents a day, up from 20, and the daily limit on deep research runs, which is the same for everyone, is now 10, up from 5. The guest allowance for designed documents, the allowance for documents with the richest artwork and the web look-up allowances are unchanged, and so is the declared free line: 3 designed documents and 3 deep research runs a day stay the floor. Each one still draws on an account's weekly allowance, or a guest's daily one, and what you see on reaching a daily limit is the same as before.

Designed documents a day, signed in

40

Deep research runs a day

10

GLM 5.2 has been retired: the company that served it stopped serving it, so it has left the model picker. Conversations that were pinned to it keep their answers and the name that wrote them; the next message asks you to pick another model. GLM 5.3, Zhipu's later release in the same family, is on the shelf in its place — text only, with the same very long context. It is new to us: we have not yet run it through our own tests, so its note on the shelf says only what it is.

On the larger models, ask for a deck in your own words — “a five-slide deck on…”, “put together a presentation for…” — and the model designs it itself; Design, under the + button, does the same there. On the free models (Llama 4 Scout, Gemma 3 27B, Mistral Small 3, Qwen 3.5 35B and GPT-OSS 120B) add Design under the + button — or tap the deck chip — and you see the plan first, then the model writes the slides; the deck is text only for now. A deck asked for in words alone on a free model is answered in the thread like any other message. Asking to change a deck you already have — “make slide 3 shorter” — is read as a change to that deck, not a request for another.

Vresk no longer chooses an engine for you. A new chat asks you to pick a model to run — the chip beside Send reads “Pick a model to run” until you do — and every answer is still labeled with the model that wrote it. Your last pick is remembered in this browser, so the next new chat starts from it. Automatic choice has been removed; older entries that mention it describe how it used to work.

The sign-in page used to offer “Send the confirmation email again” to everyone, including people who had never made an account. It now appears after a sign-in that did not work, after a confirmation link that did not, and for a guest whose new address is waiting to be confirmed. It still answers the same way whatever address you type, so it tells no one whether an address has an account here.

The composer's menu — the sparkle button beside the text box — no longer offers “Search the web”, “Research it properly” or “Run the numbers”, and a new conversation no longer shows chips for them. Ask in your own words instead — “search the web for…”, “research … thoroughly”, “what is 18% of 4,350” — and Vresk does that, as it already did without a switch. Deep research keeps running the same way, from what you ask. Designing a deck or document stays in that menu (the + button on a phone).

Free accounts now have a weekly allowance. It covers everyday answers on the everyday engines, and it resets every seven days on the weekday your account was made, so your week is your own rather than the calendar's. When it is used up, answers pause until it resets; nothing you have made is lost. The most expensive engines and long thinking are part of Plus and are not included. The daily limits on designed documents and research runs are unchanged, and like everything else on a free account they now draw on the weekly allowance. Guests keep their daily allowance exactly as before.

There is now a Sign in link on the home page and in the app, which sounds small and was not: if you already had an account, nothing anywhere in the product took you to a sign-in form, so returning visitors landed as new guests with their own history invisible behind an account they could not reach. The sign-in and create-account screens have also been rebuilt — they are separate pages now, each with one heading that says what it is for, and they carry the same mark, type and colours as the rest of Vresk instead of looking like a form someone bolted on.

When you create an account you now confirm you are 18 or older on the form itself. It was supposed to work this way since August and did not, so every account created since then met the confirmation again as a wall in front of its first message. Guests still answer it once before their first message, and nobody is asked twice.

Setting a new password now works only from a fresh reset link — before, any signed-in session could set one, including a guest session you never asked for. Finishing a reset still signs out every other device, as it always has. Sign-in failures now give one answer whether the password was wrong or the address was never confirmed, because two different answers told anyone asking whether an address has an account here. If you are waiting on a confirmation email there is a button to send it again, and it is always there rather than appearing only for accounts that exist.

Update, 5 October 2026: The resend button now shows when a confirmation can be waiting, not to everyone — see the entry of that date above.

If your account had not confirmed 18-or-older, the tasks page retried a refusal eight times and then told you to reload the page — which could never have worked, because the answer was not on that page. It now shows the one-tap confirmation itself, and your tasks load as soon as you answer it.

Guests can already have one answer being written at a time. Until now that one answer was counted per network address rather than per person, so several first-time guests sharing a school, office or mobile network took turns without knowing it: one got an answer and the rest were asked to finish the answer someone else was reading. The limit is unchanged and the allowances are unchanged; it is now counted per guest, so your one answer is your own. Signed-in accounts were always counted per person and are unaffected.

Every account gets 3 designed documents and 3 deep research runs a day, free, permanently. We are declaring it here so that it is a commitment rather than a setting. In practice the product currently allows more than this on some accounts; that headroom can move, but the numbers above are the floor and lowering them would appear here first.

Designed documents a day, every account (the promised floor)

3

Deep research runs a day, every account (the promised floor)

3

The limits the product enforces right now. Guests get a tighter bucket on the most expensive path, because that is the one worth protecting from abuse; signing in raises it. Web look-ups are counted separately from designed documents, and running out of the richest artwork means a document is still built rather than refused.

Designed documents a day, as a guest

5

Designed documents a day, signed in

20

Documents a day with the richest artwork

5

Deep research runs a day

5

Web look-ups a day, as a guest

20

Web look-ups a day, signed in

100

Completed answers now carry the name of the engine that wrote them, with a note on where that model was built and whether Vresk chose it or you pinned it. You can pin an engine to a conversation from the same place, and go back to automatic in one tap.

Update, 5 October 2026: Automatic choice has since been removed: you pick the model, and the label still names it.

A new privacy page: download everything we hold that is yours as one file, delete everything you have made while keeping your account, or delete the account outright. Deletion is real deletion, not hiding. The page also sets out, in plain language, where your conversations are processed and what we keep.

For the record, the standing instruction layers an answer is composed with today: a safety layer at version 1, a general-chat layer at version 5, and a homework-maths layer at version 1. There is no separate personality layer yet — when one ships it will appear here.