TL;DR: DeepSeek got dearer, chats acquired static links, and “For work and life” learned to notice when the train, plane and city in one question do not quite agree. Nothing dramatic — just another week.
Last week, there were two billion tokens a day and a cautious note that this was not the limit. Our colleagues at DeepSeek, over in the Middle Kingdom, appear to have read it as a commercial proposal.
Seyka is not one model. Under the bonnet sit DeepSeek, GLM and MiniMax — mostly Chinese — with GigaChat and YandexGPT also used for Russian-language work. Different models get the jobs they are best at. Geography here is not a philosophical category; it is a collection of payment details for keeping the bills paid.
DeepSeek does most of the work. Its prices rose substantially this weekend: depending on the time of day, token prices are now three to six times higher — cheaper at night, dearer in the morning. This is not a catastrophe. Catastrophes are generally harder to calculate.
The price rise did not catch our beloved moggy unawares. Recent work has cut redundant context, made tool results lighter, batched search requests, and served more repeated reading from cache. Models are selected more carefully for each task, and expensive work is no longer handed out where a fast model will do. All of this is already saving a great deal. Two billion of anything per day does, however, remain two billion. Nearly two and a half now, apparently.
The SEO chat therefore now has two general usage windows: five hours and one week, with a separate weekly allowance for Pro-tier models in any mode. Their current state is on the Usage page, with meters, automatic reset times and a breakdown of which conversations use the most.
The limits are deliberately generous. Most people are likely to meet them only by visiting that page. Pro is a little tighter: the underlying price of deep runs has risen rather further than reasonable optimism. If the weekly Pro allowance runs out, work can continue in Lite. No grand lockout, just a cheaper engine.
The system will continue to be measured and optimised, but subscriptions may now arrive earlier than intended: unfortunately, free forever is not among the more sustainable business models in AI, particularly if you are not Sberbank. The upside is that the product gets the resources to grow into something more substantial and genuinely interesting. Everyone who signed up or signs up during early access will receive a generous free bonus when subscriptions launch.
Share has always been a simple button and, for some reason, a rarely used one. It now does exactly one thing: creates a static snapshot of the conversation. The snapshot does not change with the chat, does not expose the research machinery, and lives for 30 days. It can be revoked sooner; deleting the chat removes it too.
The link is useful in WhatsApp, Telegram and email. It is particularly good with your favourite coding agent — Codex, Claude Code, OpenCode, or whichever one appears by the time this paragraph has been read.
The recipe is short:
The agent gets the research, reasoning and sources instead of fourteen screenshots and the note “there was an important table somewhere in there”. Magic follows 🥁, but we’ll leave that for you to experience. Going a step further, Seyka AI also has an MCP integration — see For developers.
Everything stays on your terms: the research is yours, the code is yours, and so is the decision to open the link to an agent. The address has a random, extremely hard-to-guess identifier, but it is not a password. Anyone with the link can open the page. Passwords, API keys and crypto wallet seed phrases do not belong in a public chat. Especially seed phrases: calling the loss “decentralised” will not make it anyone else's problem.
“For work and life” — or “Seyka without SEO”, as it is known around here — has had a substantial rebuild. It searches faster, reads more useful text, distinguishes exact facts from approximate promises more carefully, presents sources more cleanly, and does not bury disagreements between them beneath a confident sentence. Confidence is an excellent quality when the checking happened first.
One test prompt went roughly like this:
What time should I board the Burevestnik train to catch an Air Serbia flight from Moscow to Belgodod in the second half of September?
One line contained a train, a flight from Moscow, Air Serbia and a city called “Belgodod”. Alice decided we were going to Belgorod, complained that there were no flights, and sent us off on a sequence of suburban trains. A little more excitement than one might have wanted, admittedly, but resourceful.
Seyka treated Air Serbia as a clue rather than decorative wording, inferred that the person probably meant Belgrade 🇷🇸, then worked through the route and sensible transfer margin for a journey that exists in our version of reality. A fairly minor distinction, unless one prefers arriving where one intended.
Our answer to the so-called market leaders is best shared not with a coding agent but with people: friends, family, colleagues and whichever random stranger is equally fed up with their flight now so often failing to line up with the train — and where it does, access is not always available. Just send them a link to this update or to seyka.ru/en/chat — they’ll understand the rest.
And unlike the main chat in the first section, “For work and life” remains free, available without signing up, and unlimited. For the immediate future, at least. After a section on economics, promising eternity would be a bit rich.
Until then, don't forget to tell your friends about Seyka.
Peace and love.