TL;DR: documents in Seyka have learned to talk back — you can now edit them together with her in one window. The home page works out for itself whether you mean SEO or the weather in Tbilisi. And Lite and Profi have moved to new models: faster, smarter and several times cheaper than the old ones.
The previous update said September promised to be interesting. Thirteen days have passed. Reporting in. As usual, only the headlines; trifles like the fifty-odd merged pull requests are left out.
It is just that talking to your documents used to be a symptom, and as of this week it is a workflow.
The Documents section has so far been an archive: the assistant wrote something, the document went on a shelf, and all you could do with it was open it, copy it and carry it away. Every document now has a chat of its own — the very conversation it was born in. The text on the right, Seyka on the left. From there it works like a living editor: “cut it in half”, “make the tone drier”, “rewrite the third paragraph and leave the rest alone”. The edit lands straight in the document, the version number in the header ticks up, and you watch with a cup of whatever is usually in your cup.
You can edit by hand too — the Edit button opens an editor right where the document is. HTML pages — landing pages and longreads — are edited inside the page itself, but text only: as the hint honestly puts it, “ask the chat for layout and style”. If Seyka rewrote the document while you were editing it, nothing gets quietly merged: “Take the assistant’s version” or “Keep my edit”. Democracy. There is also New document — a blank sheet where you can ask Seyka to write something or start typing yourself. Skills such as the SEO copywriter now write into the open document instead of opening another one next door.
One more change worth saying out loud: Seyka has stopped explaining why a text is better left unwritten and started writing it. Copy for your site is her job, at whatever length. Where there is a professional risk she will say so in one sentence and carry on working, rather than the other way round. Just remember that Seyka is simply a clever text generator. We are not convinced that, should it come to it, blaming everything on a virtual cat will sound persuasive.
What is not there yet: version history — you cannot roll back, so keep a copy of the edits that matter. The document refreshes when the answer is finished, not while the model is typing. If you notice something odd, come to the support chat: actual humans answer there, and they genuinely want to know. Meanwhile we might just have a go at a website editor — stranger things have happened.
Since July Seyka has had two modes: For SEO and For work and life. Between them stood a toggle, and the visitor was expected to first determine their place in life and only then ask the question. The toggle worked flawlessly. It just turned out that nobody comes here to make that decision.
Now you can simply type on the home page. A site address, keywords, “Wordstat” — that is SEO. “What will the weather be in Tbilisi next week?” is life, and Seyka will not offer to check Tbilisi’s indexation. Borderline cases go to a separate small model — in a couple of seconds and before the first word of the answer, so there is no “please wait while we work out who you are” screen. The toggle stays for those who like choosing for themselves. We are toying with a third mode — image generation, straight into your documents — but are not sure anyone would use it. What do you think?
The mode is fixed to the chat by the first substantive message and does not change afterwards; “hi” decides nothing — Seyka answers briefly and waits for the real question. If the weather in Tbilisi is looking better and the air fresher, but the queue at the Verkhny Lars border crossing has just come to mind — you can find out without leaving a site audit, with the /btw command: one question in the other mode, badged “By the way”. When in doubt the home page leans towards SEO: that is, after all, what the site is about. The thresholds are set on first judgement and will be tuned against real requests; first judgement, as is well known, is occasionally wrong too.
Both modes remain free for now, in early access, and the same logic runs in Telegram at @seyka_ai_bot — so “For work and life” has reached the messenger too, as promised. Though you may only manage to open it somewhere around Pasanauri.
In August, tokens learned to cost money. In September they learned to cost a little less. Not our doing, but we have no regrets and are simply enjoying the result.
DeepSeek released V4.1 Flash. Our Pi agent bots, with the beloved moggy’s blessing, ran it through hundreds of real requests in record time — and it turned out better than the “old” V4 Pro and, which is hard not to enjoy, cheaper. DeepSeek itself appears to agree: from 14 September the old Pro will quietly turn into Flash. The same story with GLM 5.3 Flash: it has replaced the flagship GLM 5.3 on the Profi tier, costs less and is none the worse for it.
Hence a change in how the tiers work. Lite and Profi used to be two different models. Now they are one model with two temperaments: Lite gives quick answers without long reasoning; Profi is the same model thinking longer and answering more precisely. The separate Reasoning switch has gone: depth of thought is now a property of the tier, not one more knob to remember. Profi digs deeper, thinks longer and eats more tokens — but it also finds things you would never expect to find anywhere. Except, of course, the banknote your relatives hid half a century ago in a book about where the next generation of Soviet people was going to live.
Ultra — our flagship agent mode with a supervisor and subagents — has become a switch on top of the tier rather than a third item in the list. Where there used to be only Ultra-Profi, Ultra now comes in Lite too — and an audit in that mode does not draw on the weekly Profi allowance. The one exception is GigaChat: its tiers remain two models, and it has no Ultra — nobody has yet taught this “trust me bro” market leader to think and call tools at the same time, and we checked twice. As the folk wisdom goes: take it up with whichever branch issued your tokens.
The upshot: cheaper, faster and smarter in a single update. That happens rarely, and it is unlikely to become a habit. The August limits stay where they are, but at the new prices they drain noticeably slower.
Tell your friends about Seyka — you can now do it without explaining which mode they should go to.
Until then — peace and love.